kernel/sched/auto_group.c

   1 #include "sched.h"
   2
   3 #include <linux/proc_fs.h>
   4 #include <linux/seq_file.h>
   5 #include <linux/kallsyms.h>
   6 #include <linux/utsname.h>
   7 #include <linux/security.h>
   8 #include <linux/export.h>
   9
  10 unsigned int __read_mostly sysctl_sched_autogroup_enabled = 1;
  11 static struct autogroup autogroup_default;
  12 static atomic_t autogroup_seq_nr;
  13
  14 void __init autogroup_init(struct task_struct *init_task)
  15 {
  16         autogroup_default.tg = &root_task_group;
  17         kref_init(&autogroup_default.kref);
  18         init_rwsem(&autogroup_default.lock);
  19         init_task->signal->autogroup = &autogroup_default;
  20 }
  21
  22 void autogroup_free(struct task_group *tg)
  23 {
  24         kfree(tg->autogroup);
  25 }
  26
  27 static inline void autogroup_destroy(struct kref *kref)
  28 {
  29         struct autogroup *ag = container_of(kref, struct autogroup, kref);
  30
  31 #ifdef CONFIG_RT_GROUP_SCHED
  32         /* We've redirected RT tasks to the root task group... */
  33         ag->tg->rt_se = NULL;
  34         ag->tg->rt_rq = NULL;
  35 #endif
  36         sched_offline_group(ag->tg);
  37         sched_destroy_group(ag->tg);
  38 }
  39
  40 static inline void autogroup_kref_put(struct autogroup *ag)
  41 {
  42         kref_put(&ag->kref, autogroup_destroy);
  43 }
  44
  45 static inline struct autogroup *autogroup_kref_get(struct autogroup *ag)
  46 {
  47         kref_get(&ag->kref);
  48         return ag;
  49 }
  50
  51 static inline struct autogroup *autogroup_task_get(struct task_struct *p)
  52 {
  53         struct autogroup *ag;
  54         unsigned long flags;
  55
  56         if (!lock_task_sighand(p, &flags))
  57                 return autogroup_kref_get(&autogroup_default);
  58
  59         ag = autogroup_kref_get(p->signal->autogroup);
  60         unlock_task_sighand(p, &flags);
  61
  62         return ag;
  63 }
  64
  65 static inline struct autogroup *autogroup_create(void)
  66 {
  67         struct autogroup *ag = kzalloc(sizeof(*ag), GFP_KERNEL);
  68         struct task_group *tg;
  69
  70         if (!ag)
  71                 goto out_fail;
  72
  73         tg = sched_create_group(&root_task_group);
  74
  75         if (IS_ERR(tg))
  76                 goto out_free;
  77
  78         kref_init(&ag->kref);
  79         init_rwsem(&ag->lock);
  80         ag->id = atomic_inc_return(&autogroup_seq_nr);
  81         ag->tg = tg;
  82 #ifdef CONFIG_RT_GROUP_SCHED
  83         /*
  84          * Autogroup RT tasks are redirected to the root task group
  85          * so we don't have to move tasks around upon policy change,
  86          * or flail around trying to allocate bandwidth on the fly.
  87          * A bandwidth exception in __sched_setscheduler() allows
  88          * the policy change to proceed.
  89          */
  90         free_rt_sched_group(tg);
  91         tg->rt_se = root_task_group.rt_se;
  92         tg->rt_rq = root_task_group.rt_rq;
  93 #endif
  94         tg->autogroup = ag;
  95
  96         sched_online_group(tg, &root_task_group);
  97         return ag;
  98
  99 out_free:
 100         kfree(ag);
 101 out_fail:
 102         if (printk_ratelimit()) {
 103                 printk(KERN_WARNING "autogroup_create: %s failure.\n",
 104                         ag ? "sched_create_group()" : "kmalloc()");
 105         }
 106
 107         return autogroup_kref_get(&autogroup_default);
 108 }
 109
 110 bool task_wants_autogroup(struct task_struct *p, struct task_group *tg)
 111 {
 112         if (tg != &root_task_group)
 113                 return false;
 114         /*
 115          * If we race with autogroup_move_group() the caller can use the old
 116          * value of signal->autogroup but in this case sched_move_task() will
 117          * be called again before autogroup_kref_put().
 118          */
 119         return true;
 120 }
 121
 122 static void
 123 autogroup_move_group(struct task_struct *p, struct autogroup *ag)
 124 {
 125         struct autogroup *prev;
 126         struct task_struct *t;
 127         unsigned long flags;
 128
 129         BUG_ON(!lock_task_sighand(p, &flags));
 130
 131         prev = p->signal->autogroup;
 132         if (prev == ag) {
 133                 unlock_task_sighand(p, &flags);
 134                 return;
 135         }
 136
 137         p->signal->autogroup = autogroup_kref_get(ag);
 138         /*
 139          * We can't avoid sched_move_task() after we changed signal->autogroup,
 140          * this process can already run with task_group() == prev->tg or we can
 141          * race with cgroup code which can read autogroup = prev under rq->lock.
 142          * In the latter case for_each_thread() can not miss a migrating thread,
 143          * cpu_cgroup_attach() must not be possible after cgroup_exit() and it
 144          * can't be removed from thread list, we hold ->siglock.
 145          */
 146         for_each_thread(p, t)
 147                 sched_move_task(t);
 148
 149         unlock_task_sighand(p, &flags);
 150         autogroup_kref_put(prev);
 151 }
 152
 153 /* Allocates GFP_KERNEL, cannot be called under any spinlock */
 154 void sched_autogroup_create_attach(struct task_struct *p)
 155 {
 156         struct autogroup *ag = autogroup_create();
 157
 158         autogroup_move_group(p, ag);
 159         /* drop extra reference added by autogroup_create() */
 160         autogroup_kref_put(ag);
 161 }
 162 EXPORT_SYMBOL(sched_autogroup_create_attach);
 163
 164 /* Cannot be called under siglock.  Currently has no users */
 165 void sched_autogroup_detach(struct task_struct *p)
 166 {
 167         autogroup_move_group(p, &autogroup_default);
 168 }
 169 EXPORT_SYMBOL(sched_autogroup_detach);
 170
 171 void sched_autogroup_fork(struct signal_struct *sig)
 172 {
 173         sig->autogroup = autogroup_task_get(current);
 174 }
 175
 176 void sched_autogroup_exit(struct signal_struct *sig)
 177 {
 178         autogroup_kref_put(sig->autogroup);
 179 }
 180
 181 static int __init setup_autogroup(char *str)
 182 {
 183         sysctl_sched_autogroup_enabled = 0;
 184
 185         return 1;
 186 }
 187
 188 __setup("noautogroup", setup_autogroup);
 189
 190 #ifdef CONFIG_PROC_FS
 191
 192 int proc_sched_autogroup_set_nice(struct task_struct *p, int nice)
 193 {
 194         static unsigned long next = INITIAL_JIFFIES;
 195         struct autogroup *ag;
 196         int err;
 197
 198         if (nice < MIN_NICE || nice > MAX_NICE)
 199                 return -EINVAL;
 200
 201         err = security_task_setnice(current, nice);
 202         if (err)
 203                 return err;
 204
 205         if (nice < 0 && !can_nice(current, nice))
 206                 return -EPERM;
 207
 208         /* this is a heavy operation taking global locks.. */
 209         if (!capable(CAP_SYS_ADMIN) && time_before(jiffies, next))
 210                 return -EAGAIN;
 211
 212         next = HZ / 10 + jiffies;
 213         ag = autogroup_task_get(p);
 214
 215         down_write(&ag->lock);
 216         err = sched_group_set_shares(ag->tg, prio_to_weight[nice + 20]);
 217         if (!err)
 218                 ag->nice = nice;
 219         up_write(&ag->lock);
 220
 221         autogroup_kref_put(ag);
 222
 223         return err;
 224 }
 225
 226 void proc_sched_autogroup_show_task(struct task_struct *p, struct seq_file *m)
 227 {
 228         struct autogroup *ag = autogroup_task_get(p);
 229
 230         if (!task_group_is_autogroup(ag->tg))
 231                 goto out;
 232
 233         down_read(&ag->lock);
 234         seq_printf(m, "/autogroup-%ld nice %d\n", ag->id, ag->nice);
 235         up_read(&ag->lock);
 236
 237 out:
 238         autogroup_kref_put(ag);
 239 }
 240 #endif /* CONFIG_PROC_FS */
 241
 242 #ifdef CONFIG_SCHED_DEBUG
 243 int autogroup_path(struct task_group *tg, char *buf, int buflen)
 244 {
 245         if (!task_group_is_autogroup(tg))
 246                 return 0;
 247
 248         return snprintf(buf, buflen, "%s-%ld", "/autogroup", tg->autogroup->id);
 249 }
 250 #endif /* CONFIG_SCHED_DEBUG */