kernel/sched/autogroup.c

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