kernel/sched/autogroup.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /*
   4  * Auto-group scheduling implementation:
   5  */
   6
   7 unsigned int __read_mostly sysctl_sched_autogroup_enabled = 1;
   8 static struct autogroup autogroup_default;
   9 static atomic_t autogroup_seq_nr;
  10
  11 #ifdef CONFIG_SYSCTL
  12 static struct ctl_table sched_autogroup_sysctls[] = {
  13         {
  14                 .procname       = "sched_autogroup_enabled",
  15                 .data           = &sysctl_sched_autogroup_enabled,
  16                 .maxlen         = sizeof(unsigned int),
  17                 .mode           = 0644,
  18                 .proc_handler   = proc_dointvec_minmax,
  19                 .extra1         = SYSCTL_ZERO,
  20                 .extra2         = SYSCTL_ONE,
  21         },
  22 };
  23
  24 static void __init sched_autogroup_sysctl_init(void)
  25 {
  26         register_sysctl_init("kernel", sched_autogroup_sysctls);
  27 }
  28 #else
  29 #define sched_autogroup_sysctl_init() do { } while (0)
  30 #endif
  31
  32 void __init autogroup_init(struct task_struct *init_task)
  33 {
  34         autogroup_default.tg = &root_task_group;
  35         kref_init(&autogroup_default.kref);
  36         init_rwsem(&autogroup_default.lock);
  37         init_task->signal->autogroup = &autogroup_default;
  38         sched_autogroup_sysctl_init();
  39 }
  40
  41 void autogroup_free(struct task_group *tg)
  42 {
  43         kfree(tg->autogroup);
  44 }
  45
  46 static inline void autogroup_destroy(struct kref *kref)
  47 {
  48         struct autogroup *ag = container_of(kref, struct autogroup, kref);
  49
  50 #ifdef CONFIG_RT_GROUP_SCHED
  51         /* We've redirected RT tasks to the root task group... */
  52         ag->tg->rt_se = NULL;
  53         ag->tg->rt_rq = NULL;
  54 #endif
  55         sched_release_group(ag->tg);
  56         sched_destroy_group(ag->tg);
  57 }
  58
  59 static inline void autogroup_kref_put(struct autogroup *ag)
  60 {
  61         kref_put(&ag->kref, autogroup_destroy);
  62 }
  63
  64 static inline struct autogroup *autogroup_kref_get(struct autogroup *ag)
  65 {
  66         kref_get(&ag->kref);
  67         return ag;
  68 }
  69
  70 static inline struct autogroup *autogroup_task_get(struct task_struct *p)
  71 {
  72         struct autogroup *ag;
  73         unsigned long flags;
  74
  75         if (!lock_task_sighand(p, &flags))
  76                 return autogroup_kref_get(&autogroup_default);
  77
  78         ag = autogroup_kref_get(p->signal->autogroup);
  79         unlock_task_sighand(p, &flags);
  80
  81         return ag;
  82 }
  83
  84 static inline struct autogroup *autogroup_create(void)
  85 {
  86         struct autogroup *ag = kzalloc(sizeof(*ag), GFP_KERNEL);
  87         struct task_group *tg;
  88
  89         if (!ag)
  90                 goto out_fail;
  91
  92         tg = sched_create_group(&root_task_group);
  93         if (IS_ERR(tg))
  94                 goto out_free;
  95
  96         kref_init(&ag->kref);
  97         init_rwsem(&ag->lock);
  98         ag->id = atomic_inc_return(&autogroup_seq_nr);
  99         ag->tg = tg;
 100 #ifdef CONFIG_RT_GROUP_SCHED
 101         /*
 102          * Autogroup RT tasks are redirected to the root task group
 103          * so we don't have to move tasks around upon policy change,
 104          * or flail around trying to allocate bandwidth on the fly.
 105          * A bandwidth exception in __sched_setscheduler() allows
 106          * the policy change to proceed.
 107          */
 108         free_rt_sched_group(tg);
 109         tg->rt_se = root_task_group.rt_se;
 110         tg->rt_rq = root_task_group.rt_rq;
 111 #endif
 112         tg->autogroup = ag;
 113
 114         sched_online_group(tg, &root_task_group);
 115         return ag;
 116
 117 out_free:
 118         kfree(ag);
 119 out_fail:
 120         if (printk_ratelimit()) {
 121                 printk(KERN_WARNING "autogroup_create: %s failure.\n",
 122                         ag ? "sched_create_group()" : "kzalloc()");
 123         }
 124
 125         return autogroup_kref_get(&autogroup_default);
 126 }
 127
 128 bool task_wants_autogroup(struct task_struct *p, struct task_group *tg)
 129 {
 130         if (tg != &root_task_group)
 131                 return false;
 132         /*
 133          * If we race with autogroup_move_group() the caller can use the old
 134          * value of signal->autogroup but in this case sched_move_task() will
 135          * be called again before autogroup_kref_put().
 136          *
 137          * However, there is no way sched_autogroup_exit_task() could tell us
 138          * to avoid autogroup->tg, so we abuse PF_EXITING flag for this case.
 139          */
 140         if (p->flags & PF_EXITING)
 141                 return false;
 142
 143         return true;
 144 }
 145
 146 void sched_autogroup_exit_task(struct task_struct *p)
 147 {
 148         /*
 149          * We are going to call exit_notify() and autogroup_move_group() can't
 150          * see this thread after that: we can no longer use signal->autogroup.
 151          * See the PF_EXITING check in task_wants_autogroup().
 152          */
 153         sched_move_task(p);
 154 }
 155
 156 static void
 157 autogroup_move_group(struct task_struct *p, struct autogroup *ag)
 158 {
 159         struct autogroup *prev;
 160         struct task_struct *t;
 161         unsigned long flags;
 162
 163         if (WARN_ON_ONCE(!lock_task_sighand(p, &flags)))
 164                 return;
 165
 166         prev = p->signal->autogroup;
 167         if (prev == ag) {
 168                 unlock_task_sighand(p, &flags);
 169                 return;
 170         }
 171
 172         p->signal->autogroup = autogroup_kref_get(ag);
 173         /*
 174          * We can't avoid sched_move_task() after we changed signal->autogroup,
 175          * this process can already run with task_group() == prev->tg or we can
 176          * race with cgroup code which can read autogroup = prev under rq->lock.
 177          * In the latter case for_each_thread() can not miss a migrating thread,
 178          * cpu_cgroup_attach() must not be possible after cgroup_exit() and it
 179          * can't be removed from thread list, we hold ->siglock.
 180          *
 181          * If an exiting thread was already removed from thread list we rely on
 182          * sched_autogroup_exit_task().
 183          */
 184         for_each_thread(p, t)
 185                 sched_move_task(t);
 186
 187         unlock_task_sighand(p, &flags);
 188         autogroup_kref_put(prev);
 189 }
 190
 191 /* Allocates GFP_KERNEL, cannot be called under any spinlock: */
 192 void sched_autogroup_create_attach(struct task_struct *p)
 193 {
 194         struct autogroup *ag = autogroup_create();
 195
 196         autogroup_move_group(p, ag);
 197
 198         /* Drop extra reference added by autogroup_create(): */
 199         autogroup_kref_put(ag);
 200 }
 201 EXPORT_SYMBOL(sched_autogroup_create_attach);
 202
 203 /* Cannot be called under siglock. Currently has no users: */
 204 void sched_autogroup_detach(struct task_struct *p)
 205 {
 206         autogroup_move_group(p, &autogroup_default);
 207 }
 208 EXPORT_SYMBOL(sched_autogroup_detach);
 209
 210 void sched_autogroup_fork(struct signal_struct *sig)
 211 {
 212         sig->autogroup = autogroup_task_get(current);
 213 }
 214
 215 void sched_autogroup_exit(struct signal_struct *sig)
 216 {
 217         autogroup_kref_put(sig->autogroup);
 218 }
 219
 220 static int __init setup_autogroup(char *str)
 221 {
 222         sysctl_sched_autogroup_enabled = 0;
 223
 224         return 1;
 225 }
 226 __setup("noautogroup", setup_autogroup);
 227
 228 #ifdef CONFIG_PROC_FS
 229
 230 int proc_sched_autogroup_set_nice(struct task_struct *p, int nice)
 231 {
 232         static unsigned long next = INITIAL_JIFFIES;
 233         struct autogroup *ag;
 234         unsigned long shares;
 235         int err, idx;
 236
 237         if (nice < MIN_NICE || nice > MAX_NICE)
 238                 return -EINVAL;
 239
 240         err = security_task_setnice(current, nice);
 241         if (err)
 242                 return err;
 243
 244         if (nice < 0 && !can_nice(current, nice))
 245                 return -EPERM;
 246
 247         /* This is a heavy operation, taking global locks.. */
 248         if (!capable(CAP_SYS_ADMIN) && time_before(jiffies, next))
 249                 return -EAGAIN;
 250
 251         next = HZ / 10 + jiffies;
 252         ag = autogroup_task_get(p);
 253
 254         idx = array_index_nospec(nice + 20, 40);
 255         shares = scale_load(sched_prio_to_weight[idx]);
 256
 257         down_write(&ag->lock);
 258         err = sched_group_set_shares(ag->tg, shares);
 259         if (!err)
 260                 ag->nice = nice;
 261         up_write(&ag->lock);
 262
 263         autogroup_kref_put(ag);
 264
 265         return err;
 266 }
 267
 268 void proc_sched_autogroup_show_task(struct task_struct *p, struct seq_file *m)
 269 {
 270         struct autogroup *ag = autogroup_task_get(p);
 271
 272         if (!task_group_is_autogroup(ag->tg))
 273                 goto out;
 274
 275         down_read(&ag->lock);
 276         seq_printf(m, "/autogroup-%ld nice %d\n", ag->id, ag->nice);
 277         up_read(&ag->lock);
 278
 279 out:
 280         autogroup_kref_put(ag);
 281 }
 282 #endif /* CONFIG_PROC_FS */
 283
 284 int autogroup_path(struct task_group *tg, char *buf, int buflen)
 285 {
 286         if (!task_group_is_autogroup(tg))
 287                 return 0;
 288
 289         return snprintf(buf, buflen, "%s-%ld", "/autogroup", tg->autogroup->id);
 290 }