1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * taskstats.c - Export per-task statistics to userland
5 * Copyright (C) Shailabh Nagar, IBM Corp. 2006
6 * (C) Balbir Singh, IBM Corp. 2006
9 #include <linux/kernel.h>
10 #include <linux/taskstats_kern.h>
11 #include <linux/tsacct_kern.h>
12 #include <linux/delayacct.h>
13 #include <linux/cpumask.h>
14 #include <linux/percpu.h>
15 #include <linux/slab.h>
16 #include <linux/cgroupstats.h>
17 #include <linux/cgroup.h>
19 #include <linux/file.h>
20 #include <linux/pid_namespace.h>
21 #include <net/genetlink.h>
22 #include <linux/atomic.h>
23 #include <linux/sched/cputime.h>
26 * Maximum length of a cpumask that can be specified in
27 * the TASKSTATS_CMD_ATTR_REGISTER/DEREGISTER_CPUMASK attribute
29 #define TASKSTATS_CPUMASK_MAXLEN (100+6*NR_CPUS)
31 static DEFINE_PER_CPU(__u32, taskstats_seqnum);
32 static int family_registered;
33 struct kmem_cache *taskstats_cache;
35 static struct genl_family family;
37 static const struct nla_policy taskstats_cmd_get_policy[TASKSTATS_CMD_ATTR_MAX+1] = {
38 [TASKSTATS_CMD_ATTR_PID] = { .type = NLA_U32 },
39 [TASKSTATS_CMD_ATTR_TGID] = { .type = NLA_U32 },
40 [TASKSTATS_CMD_ATTR_REGISTER_CPUMASK] = { .type = NLA_STRING },
41 [TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK] = { .type = NLA_STRING },};
44 * We have to use TASKSTATS_CMD_ATTR_MAX here, it is the maxattr in the family.
45 * Make sure they are always aligned.
47 static const struct nla_policy cgroupstats_cmd_get_policy[TASKSTATS_CMD_ATTR_MAX+1] = {
48 [CGROUPSTATS_CMD_ATTR_FD] = { .type = NLA_U32 },
52 struct list_head list;
57 struct listener_list {
58 struct rw_semaphore sem;
59 struct list_head list;
61 static DEFINE_PER_CPU(struct listener_list, listener_array);
69 static int prepare_reply(struct genl_info *info, u8 cmd, struct sk_buff **skbp,
76 * If new attributes are added, please revisit this allocation
78 skb = genlmsg_new(size, GFP_KERNEL);
83 int seq = this_cpu_inc_return(taskstats_seqnum) - 1;
85 reply = genlmsg_put(skb, 0, seq, &family, 0, cmd);
87 reply = genlmsg_put_reply(skb, info, &family, 0, cmd);
98 * Send taskstats data in @skb to listener with nl_pid @pid
100 static int send_reply(struct sk_buff *skb, struct genl_info *info)
102 struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
103 void *reply = genlmsg_data(genlhdr);
105 genlmsg_end(skb, reply);
107 return genlmsg_reply(skb, info);
111 * Send taskstats data in @skb to listeners registered for @cpu's exit data
113 static void send_cpu_listeners(struct sk_buff *skb,
114 struct listener_list *listeners)
116 struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
117 struct listener *s, *tmp;
118 struct sk_buff *skb_next, *skb_cur = skb;
119 void *reply = genlmsg_data(genlhdr);
120 int rc, delcount = 0;
122 genlmsg_end(skb, reply);
125 down_read(&listeners->sem);
126 list_for_each_entry(s, &listeners->list, list) {
128 if (!list_is_last(&s->list, &listeners->list)) {
129 skb_next = skb_clone(skb_cur, GFP_KERNEL);
133 rc = genlmsg_unicast(&init_net, skb_cur, s->pid);
134 if (rc == -ECONNREFUSED) {
140 up_read(&listeners->sem);
148 /* Delete invalidated entries */
149 down_write(&listeners->sem);
150 list_for_each_entry_safe(s, tmp, &listeners->list, list) {
156 up_write(&listeners->sem);
159 static void fill_stats(struct user_namespace *user_ns,
160 struct pid_namespace *pid_ns,
161 struct task_struct *tsk, struct taskstats *stats)
163 memset(stats, 0, sizeof(*stats));
165 * Each accounting subsystem adds calls to its functions to
166 * fill in relevant parts of struct taskstsats as follows
168 * per-task-foo(stats, tsk);
171 delayacct_add_tsk(stats, tsk);
173 /* fill in basic acct fields */
174 stats->version = TASKSTATS_VERSION;
175 stats->nvcsw = tsk->nvcsw;
176 stats->nivcsw = tsk->nivcsw;
177 bacct_add_tsk(user_ns, pid_ns, stats, tsk);
179 /* fill in extended acct fields */
180 xacct_add_tsk(stats, tsk);
183 static int fill_stats_for_pid(pid_t pid, struct taskstats *stats)
185 struct task_struct *tsk;
187 tsk = find_get_task_by_vpid(pid);
190 fill_stats(current_user_ns(), task_active_pid_ns(current), tsk, stats);
191 put_task_struct(tsk);
195 static int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
197 struct task_struct *tsk, *first;
200 u64 delta, utime, stime;
204 * Add additional stats from live tasks except zombie thread group
205 * leaders who are already counted with the dead tasks
208 first = find_task_by_vpid(tgid);
210 if (!first || !lock_task_sighand(first, &flags))
213 if (first->signal->stats)
214 memcpy(stats, first->signal->stats, sizeof(*stats));
216 memset(stats, 0, sizeof(*stats));
219 start_time = ktime_get_ns();
224 * Accounting subsystem can call its functions here to
225 * fill in relevant parts of struct taskstsats as follows
227 * per-task-foo(stats, tsk);
229 delayacct_add_tsk(stats, tsk);
231 /* calculate task elapsed time in nsec */
232 delta = start_time - tsk->start_time;
233 /* Convert to micro seconds */
234 do_div(delta, NSEC_PER_USEC);
235 stats->ac_etime += delta;
237 task_cputime(tsk, &utime, &stime);
238 stats->ac_utime += div_u64(utime, NSEC_PER_USEC);
239 stats->ac_stime += div_u64(stime, NSEC_PER_USEC);
241 stats->nvcsw += tsk->nvcsw;
242 stats->nivcsw += tsk->nivcsw;
243 } while_each_thread(first, tsk);
245 unlock_task_sighand(first, &flags);
250 stats->version = TASKSTATS_VERSION;
252 * Accounting subsystems can also add calls here to modify
253 * fields of taskstats.
258 static void fill_tgid_exit(struct task_struct *tsk)
262 spin_lock_irqsave(&tsk->sighand->siglock, flags);
263 if (!tsk->signal->stats)
267 * Each accounting subsystem calls its functions here to
268 * accumalate its per-task stats for tsk, into the per-tgid structure
270 * per-task-foo(tsk->signal->stats, tsk);
272 delayacct_add_tsk(tsk->signal->stats, tsk);
274 spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
278 static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
280 struct listener_list *listeners;
281 struct listener *s, *tmp, *s2;
285 if (!cpumask_subset(mask, cpu_possible_mask))
288 if (current_user_ns() != &init_user_ns)
291 if (task_active_pid_ns(current) != &init_pid_ns)
294 if (isadd == REGISTER) {
295 for_each_cpu(cpu, mask) {
296 s = kmalloc_node(sizeof(struct listener),
297 GFP_KERNEL, cpu_to_node(cpu));
305 listeners = &per_cpu(listener_array, cpu);
306 down_write(&listeners->sem);
307 list_for_each_entry(s2, &listeners->list, list) {
308 if (s2->pid == pid && s2->valid)
311 list_add(&s->list, &listeners->list);
314 up_write(&listeners->sem);
315 kfree(s); /* nop if NULL */
320 /* Deregister or cleanup */
322 for_each_cpu(cpu, mask) {
323 listeners = &per_cpu(listener_array, cpu);
324 down_write(&listeners->sem);
325 list_for_each_entry_safe(s, tmp, &listeners->list, list) {
332 up_write(&listeners->sem);
337 static int parse(struct nlattr *na, struct cpumask *mask)
346 if (len > TASKSTATS_CPUMASK_MAXLEN)
350 data = kmalloc(len, GFP_KERNEL);
353 nla_strlcpy(data, na, len);
354 ret = cpulist_parse(data, mask);
359 static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
361 struct nlattr *na, *ret;
364 aggr = (type == TASKSTATS_TYPE_PID)
365 ? TASKSTATS_TYPE_AGGR_PID
366 : TASKSTATS_TYPE_AGGR_TGID;
368 na = nla_nest_start_noflag(skb, aggr);
372 if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
373 nla_nest_cancel(skb, na);
376 ret = nla_reserve_64bit(skb, TASKSTATS_TYPE_STATS,
377 sizeof(struct taskstats), TASKSTATS_TYPE_NULL);
379 nla_nest_cancel(skb, na);
382 nla_nest_end(skb, na);
384 return nla_data(ret);
389 static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
392 struct sk_buff *rep_skb;
393 struct cgroupstats *stats;
399 na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
403 fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
408 size = nla_total_size(sizeof(struct cgroupstats));
410 rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
415 na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
416 sizeof(struct cgroupstats));
423 stats = nla_data(na);
424 memset(stats, 0, sizeof(*stats));
426 rc = cgroupstats_build(stats, f.file->f_path.dentry);
432 rc = send_reply(rep_skb, info);
439 static int cmd_attr_register_cpumask(struct genl_info *info)
444 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
446 rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
449 rc = add_del_listener(info->snd_portid, mask, REGISTER);
451 free_cpumask_var(mask);
455 static int cmd_attr_deregister_cpumask(struct genl_info *info)
460 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
462 rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
465 rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
467 free_cpumask_var(mask);
471 static size_t taskstats_packet_size(void)
475 size = nla_total_size(sizeof(u32)) +
476 nla_total_size_64bit(sizeof(struct taskstats)) +
482 static int cmd_attr_pid(struct genl_info *info)
484 struct taskstats *stats;
485 struct sk_buff *rep_skb;
490 size = taskstats_packet_size();
492 rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
497 pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
498 stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
502 rc = fill_stats_for_pid(pid, stats);
505 return send_reply(rep_skb, info);
511 static int cmd_attr_tgid(struct genl_info *info)
513 struct taskstats *stats;
514 struct sk_buff *rep_skb;
519 size = taskstats_packet_size();
521 rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
526 tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
527 stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
531 rc = fill_stats_for_tgid(tgid, stats);
534 return send_reply(rep_skb, info);
540 static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
542 if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
543 return cmd_attr_register_cpumask(info);
544 else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
545 return cmd_attr_deregister_cpumask(info);
546 else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
547 return cmd_attr_pid(info);
548 else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
549 return cmd_attr_tgid(info);
554 static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
556 struct signal_struct *sig = tsk->signal;
557 struct taskstats *stats_new, *stats;
559 /* Pairs with smp_store_release() below. */
560 stats = smp_load_acquire(&sig->stats);
561 if (stats || thread_group_empty(tsk))
564 /* No problem if kmem_cache_zalloc() fails */
565 stats_new = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
567 spin_lock_irq(&tsk->sighand->siglock);
571 * Pairs with smp_store_release() above and order the
572 * kmem_cache_zalloc().
574 smp_store_release(&sig->stats, stats_new);
578 spin_unlock_irq(&tsk->sighand->siglock);
581 kmem_cache_free(taskstats_cache, stats_new);
586 /* Send pid data out on exit */
587 void taskstats_exit(struct task_struct *tsk, int group_dead)
590 struct listener_list *listeners;
591 struct taskstats *stats;
592 struct sk_buff *rep_skb;
596 if (!family_registered)
600 * Size includes space for nested attributes
602 size = taskstats_packet_size();
604 is_thread_group = !!taskstats_tgid_alloc(tsk);
605 if (is_thread_group) {
606 /* PID + STATS + TGID + STATS */
608 /* fill the tsk->signal->stats structure */
612 listeners = raw_cpu_ptr(&listener_array);
613 if (list_empty(&listeners->list))
616 rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
620 stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
621 task_pid_nr_ns(tsk, &init_pid_ns));
625 fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
628 * Doesn't matter if tsk is the leader or the last group member leaving
630 if (!is_thread_group || !group_dead)
633 stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
634 task_tgid_nr_ns(tsk, &init_pid_ns));
638 memcpy(stats, tsk->signal->stats, sizeof(*stats));
641 send_cpu_listeners(rep_skb, listeners);
647 static const struct genl_ops taskstats_ops[] = {
649 .cmd = TASKSTATS_CMD_GET,
650 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
651 .doit = taskstats_user_cmd,
652 /* policy enforced later */
653 .flags = GENL_ADMIN_PERM | GENL_CMD_CAP_HASPOL,
656 .cmd = CGROUPSTATS_CMD_GET,
657 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
658 .doit = cgroupstats_user_cmd,
659 /* policy enforced later */
660 .flags = GENL_CMD_CAP_HASPOL,
664 static int taskstats_pre_doit(const struct genl_ops *ops, struct sk_buff *skb,
665 struct genl_info *info)
667 const struct nla_policy *policy = NULL;
670 case TASKSTATS_CMD_GET:
671 policy = taskstats_cmd_get_policy;
673 case CGROUPSTATS_CMD_GET:
674 policy = cgroupstats_cmd_get_policy;
680 return nlmsg_validate_deprecated(info->nlhdr, GENL_HDRLEN,
681 TASKSTATS_CMD_ATTR_MAX, policy,
685 static struct genl_family family __ro_after_init = {
686 .name = TASKSTATS_GENL_NAME,
687 .version = TASKSTATS_GENL_VERSION,
688 .maxattr = TASKSTATS_CMD_ATTR_MAX,
689 .module = THIS_MODULE,
690 .ops = taskstats_ops,
691 .n_ops = ARRAY_SIZE(taskstats_ops),
692 .pre_doit = taskstats_pre_doit,
695 /* Needed early in initialization */
696 void __init taskstats_init_early(void)
700 taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
701 for_each_possible_cpu(i) {
702 INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
703 init_rwsem(&(per_cpu(listener_array, i).sem));
707 static int __init taskstats_init(void)
711 rc = genl_register_family(&family);
715 family_registered = 1;
716 pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
721 * late initcall ensures initialization of statistics collection
722 * mechanisms precedes initialization of the taskstats interface
724 late_initcall(taskstats_init);