xref: /openbmc/linux/kernel/taskstats.c (revision 239480ab)
1 /*
2  * taskstats.c - Export per-task statistics to userland
3  *
4  * Copyright (C) Shailabh Nagar, IBM Corp. 2006
5  *           (C) Balbir Singh,   IBM Corp. 2006
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License as published by
9  * the Free Software Foundation; either version 2 of the License, or
10  * (at your option) any later version.
11  *
12  * This program is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15  * GNU General Public License for more details.
16  *
17  */
18 
19 #include <linux/kernel.h>
20 #include <linux/taskstats_kern.h>
21 #include <linux/tsacct_kern.h>
22 #include <linux/delayacct.h>
23 #include <linux/cpumask.h>
24 #include <linux/percpu.h>
25 #include <linux/slab.h>
26 #include <linux/cgroupstats.h>
27 #include <linux/cgroup.h>
28 #include <linux/fs.h>
29 #include <linux/file.h>
30 #include <linux/pid_namespace.h>
31 #include <net/genetlink.h>
32 #include <linux/atomic.h>
33 #include <linux/sched/cputime.h>
34 
35 /*
36  * Maximum length of a cpumask that can be specified in
37  * the TASKSTATS_CMD_ATTR_REGISTER/DEREGISTER_CPUMASK attribute
38  */
39 #define TASKSTATS_CPUMASK_MAXLEN	(100+6*NR_CPUS)
40 
41 static DEFINE_PER_CPU(__u32, taskstats_seqnum);
42 static int family_registered;
43 struct kmem_cache *taskstats_cache;
44 
45 static struct genl_family family;
46 
47 static const struct nla_policy taskstats_cmd_get_policy[TASKSTATS_CMD_ATTR_MAX+1] = {
48 	[TASKSTATS_CMD_ATTR_PID]  = { .type = NLA_U32 },
49 	[TASKSTATS_CMD_ATTR_TGID] = { .type = NLA_U32 },
50 	[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK] = { .type = NLA_STRING },
51 	[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK] = { .type = NLA_STRING },};
52 
53 /*
54  * We have to use TASKSTATS_CMD_ATTR_MAX here, it is the maxattr in the family.
55  * Make sure they are always aligned.
56  */
57 static const struct nla_policy cgroupstats_cmd_get_policy[TASKSTATS_CMD_ATTR_MAX+1] = {
58 	[CGROUPSTATS_CMD_ATTR_FD] = { .type = NLA_U32 },
59 };
60 
61 struct listener {
62 	struct list_head list;
63 	pid_t pid;
64 	char valid;
65 };
66 
67 struct listener_list {
68 	struct rw_semaphore sem;
69 	struct list_head list;
70 };
71 static DEFINE_PER_CPU(struct listener_list, listener_array);
72 
73 enum actions {
74 	REGISTER,
75 	DEREGISTER,
76 	CPU_DONT_CARE
77 };
78 
79 static int prepare_reply(struct genl_info *info, u8 cmd, struct sk_buff **skbp,
80 				size_t size)
81 {
82 	struct sk_buff *skb;
83 	void *reply;
84 
85 	/*
86 	 * If new attributes are added, please revisit this allocation
87 	 */
88 	skb = genlmsg_new(size, GFP_KERNEL);
89 	if (!skb)
90 		return -ENOMEM;
91 
92 	if (!info) {
93 		int seq = this_cpu_inc_return(taskstats_seqnum) - 1;
94 
95 		reply = genlmsg_put(skb, 0, seq, &family, 0, cmd);
96 	} else
97 		reply = genlmsg_put_reply(skb, info, &family, 0, cmd);
98 	if (reply == NULL) {
99 		nlmsg_free(skb);
100 		return -EINVAL;
101 	}
102 
103 	*skbp = skb;
104 	return 0;
105 }
106 
107 /*
108  * Send taskstats data in @skb to listener with nl_pid @pid
109  */
110 static int send_reply(struct sk_buff *skb, struct genl_info *info)
111 {
112 	struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
113 	void *reply = genlmsg_data(genlhdr);
114 
115 	genlmsg_end(skb, reply);
116 
117 	return genlmsg_reply(skb, info);
118 }
119 
120 /*
121  * Send taskstats data in @skb to listeners registered for @cpu's exit data
122  */
123 static void send_cpu_listeners(struct sk_buff *skb,
124 					struct listener_list *listeners)
125 {
126 	struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
127 	struct listener *s, *tmp;
128 	struct sk_buff *skb_next, *skb_cur = skb;
129 	void *reply = genlmsg_data(genlhdr);
130 	int rc, delcount = 0;
131 
132 	genlmsg_end(skb, reply);
133 
134 	rc = 0;
135 	down_read(&listeners->sem);
136 	list_for_each_entry(s, &listeners->list, list) {
137 		skb_next = NULL;
138 		if (!list_is_last(&s->list, &listeners->list)) {
139 			skb_next = skb_clone(skb_cur, GFP_KERNEL);
140 			if (!skb_next)
141 				break;
142 		}
143 		rc = genlmsg_unicast(&init_net, skb_cur, s->pid);
144 		if (rc == -ECONNREFUSED) {
145 			s->valid = 0;
146 			delcount++;
147 		}
148 		skb_cur = skb_next;
149 	}
150 	up_read(&listeners->sem);
151 
152 	if (skb_cur)
153 		nlmsg_free(skb_cur);
154 
155 	if (!delcount)
156 		return;
157 
158 	/* Delete invalidated entries */
159 	down_write(&listeners->sem);
160 	list_for_each_entry_safe(s, tmp, &listeners->list, list) {
161 		if (!s->valid) {
162 			list_del(&s->list);
163 			kfree(s);
164 		}
165 	}
166 	up_write(&listeners->sem);
167 }
168 
169 static void fill_stats(struct user_namespace *user_ns,
170 		       struct pid_namespace *pid_ns,
171 		       struct task_struct *tsk, struct taskstats *stats)
172 {
173 	memset(stats, 0, sizeof(*stats));
174 	/*
175 	 * Each accounting subsystem adds calls to its functions to
176 	 * fill in relevant parts of struct taskstsats as follows
177 	 *
178 	 *	per-task-foo(stats, tsk);
179 	 */
180 
181 	delayacct_add_tsk(stats, tsk);
182 
183 	/* fill in basic acct fields */
184 	stats->version = TASKSTATS_VERSION;
185 	stats->nvcsw = tsk->nvcsw;
186 	stats->nivcsw = tsk->nivcsw;
187 	bacct_add_tsk(user_ns, pid_ns, stats, tsk);
188 
189 	/* fill in extended acct fields */
190 	xacct_add_tsk(stats, tsk);
191 }
192 
193 static int fill_stats_for_pid(pid_t pid, struct taskstats *stats)
194 {
195 	struct task_struct *tsk;
196 
197 	rcu_read_lock();
198 	tsk = find_task_by_vpid(pid);
199 	if (tsk)
200 		get_task_struct(tsk);
201 	rcu_read_unlock();
202 	if (!tsk)
203 		return -ESRCH;
204 	fill_stats(current_user_ns(), task_active_pid_ns(current), tsk, stats);
205 	put_task_struct(tsk);
206 	return 0;
207 }
208 
209 static int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
210 {
211 	struct task_struct *tsk, *first;
212 	unsigned long flags;
213 	int rc = -ESRCH;
214 	u64 delta, utime, stime;
215 	u64 start_time;
216 
217 	/*
218 	 * Add additional stats from live tasks except zombie thread group
219 	 * leaders who are already counted with the dead tasks
220 	 */
221 	rcu_read_lock();
222 	first = find_task_by_vpid(tgid);
223 
224 	if (!first || !lock_task_sighand(first, &flags))
225 		goto out;
226 
227 	if (first->signal->stats)
228 		memcpy(stats, first->signal->stats, sizeof(*stats));
229 	else
230 		memset(stats, 0, sizeof(*stats));
231 
232 	tsk = first;
233 	start_time = ktime_get_ns();
234 	do {
235 		if (tsk->exit_state)
236 			continue;
237 		/*
238 		 * Accounting subsystem can call its functions here to
239 		 * fill in relevant parts of struct taskstsats as follows
240 		 *
241 		 *	per-task-foo(stats, tsk);
242 		 */
243 		delayacct_add_tsk(stats, tsk);
244 
245 		/* calculate task elapsed time in nsec */
246 		delta = start_time - tsk->start_time;
247 		/* Convert to micro seconds */
248 		do_div(delta, NSEC_PER_USEC);
249 		stats->ac_etime += delta;
250 
251 		task_cputime(tsk, &utime, &stime);
252 		stats->ac_utime += div_u64(utime, NSEC_PER_USEC);
253 		stats->ac_stime += div_u64(stime, NSEC_PER_USEC);
254 
255 		stats->nvcsw += tsk->nvcsw;
256 		stats->nivcsw += tsk->nivcsw;
257 	} while_each_thread(first, tsk);
258 
259 	unlock_task_sighand(first, &flags);
260 	rc = 0;
261 out:
262 	rcu_read_unlock();
263 
264 	stats->version = TASKSTATS_VERSION;
265 	/*
266 	 * Accounting subsystems can also add calls here to modify
267 	 * fields of taskstats.
268 	 */
269 	return rc;
270 }
271 
272 static void fill_tgid_exit(struct task_struct *tsk)
273 {
274 	unsigned long flags;
275 
276 	spin_lock_irqsave(&tsk->sighand->siglock, flags);
277 	if (!tsk->signal->stats)
278 		goto ret;
279 
280 	/*
281 	 * Each accounting subsystem calls its functions here to
282 	 * accumalate its per-task stats for tsk, into the per-tgid structure
283 	 *
284 	 *	per-task-foo(tsk->signal->stats, tsk);
285 	 */
286 	delayacct_add_tsk(tsk->signal->stats, tsk);
287 ret:
288 	spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
289 	return;
290 }
291 
292 static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
293 {
294 	struct listener_list *listeners;
295 	struct listener *s, *tmp, *s2;
296 	unsigned int cpu;
297 	int ret = 0;
298 
299 	if (!cpumask_subset(mask, cpu_possible_mask))
300 		return -EINVAL;
301 
302 	if (current_user_ns() != &init_user_ns)
303 		return -EINVAL;
304 
305 	if (task_active_pid_ns(current) != &init_pid_ns)
306 		return -EINVAL;
307 
308 	if (isadd == REGISTER) {
309 		for_each_cpu(cpu, mask) {
310 			s = kmalloc_node(sizeof(struct listener),
311 					GFP_KERNEL, cpu_to_node(cpu));
312 			if (!s) {
313 				ret = -ENOMEM;
314 				goto cleanup;
315 			}
316 			s->pid = pid;
317 			s->valid = 1;
318 
319 			listeners = &per_cpu(listener_array, cpu);
320 			down_write(&listeners->sem);
321 			list_for_each_entry(s2, &listeners->list, list) {
322 				if (s2->pid == pid && s2->valid)
323 					goto exists;
324 			}
325 			list_add(&s->list, &listeners->list);
326 			s = NULL;
327 exists:
328 			up_write(&listeners->sem);
329 			kfree(s); /* nop if NULL */
330 		}
331 		return 0;
332 	}
333 
334 	/* Deregister or cleanup */
335 cleanup:
336 	for_each_cpu(cpu, mask) {
337 		listeners = &per_cpu(listener_array, cpu);
338 		down_write(&listeners->sem);
339 		list_for_each_entry_safe(s, tmp, &listeners->list, list) {
340 			if (s->pid == pid) {
341 				list_del(&s->list);
342 				kfree(s);
343 				break;
344 			}
345 		}
346 		up_write(&listeners->sem);
347 	}
348 	return ret;
349 }
350 
351 static int parse(struct nlattr *na, struct cpumask *mask)
352 {
353 	char *data;
354 	int len;
355 	int ret;
356 
357 	if (na == NULL)
358 		return 1;
359 	len = nla_len(na);
360 	if (len > TASKSTATS_CPUMASK_MAXLEN)
361 		return -E2BIG;
362 	if (len < 1)
363 		return -EINVAL;
364 	data = kmalloc(len, GFP_KERNEL);
365 	if (!data)
366 		return -ENOMEM;
367 	nla_strlcpy(data, na, len);
368 	ret = cpulist_parse(data, mask);
369 	kfree(data);
370 	return ret;
371 }
372 
373 static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
374 {
375 	struct nlattr *na, *ret;
376 	int aggr;
377 
378 	aggr = (type == TASKSTATS_TYPE_PID)
379 			? TASKSTATS_TYPE_AGGR_PID
380 			: TASKSTATS_TYPE_AGGR_TGID;
381 
382 	na = nla_nest_start(skb, aggr);
383 	if (!na)
384 		goto err;
385 
386 	if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
387 		nla_nest_cancel(skb, na);
388 		goto err;
389 	}
390 	ret = nla_reserve_64bit(skb, TASKSTATS_TYPE_STATS,
391 				sizeof(struct taskstats), TASKSTATS_TYPE_NULL);
392 	if (!ret) {
393 		nla_nest_cancel(skb, na);
394 		goto err;
395 	}
396 	nla_nest_end(skb, na);
397 
398 	return nla_data(ret);
399 err:
400 	return NULL;
401 }
402 
403 static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
404 {
405 	int rc = 0;
406 	struct sk_buff *rep_skb;
407 	struct cgroupstats *stats;
408 	struct nlattr *na;
409 	size_t size;
410 	u32 fd;
411 	struct fd f;
412 
413 	na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
414 	if (!na)
415 		return -EINVAL;
416 
417 	fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
418 	f = fdget(fd);
419 	if (!f.file)
420 		return 0;
421 
422 	size = nla_total_size(sizeof(struct cgroupstats));
423 
424 	rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
425 				size);
426 	if (rc < 0)
427 		goto err;
428 
429 	na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
430 				sizeof(struct cgroupstats));
431 	if (na == NULL) {
432 		nlmsg_free(rep_skb);
433 		rc = -EMSGSIZE;
434 		goto err;
435 	}
436 
437 	stats = nla_data(na);
438 	memset(stats, 0, sizeof(*stats));
439 
440 	rc = cgroupstats_build(stats, f.file->f_path.dentry);
441 	if (rc < 0) {
442 		nlmsg_free(rep_skb);
443 		goto err;
444 	}
445 
446 	rc = send_reply(rep_skb, info);
447 
448 err:
449 	fdput(f);
450 	return rc;
451 }
452 
453 static int cmd_attr_register_cpumask(struct genl_info *info)
454 {
455 	cpumask_var_t mask;
456 	int rc;
457 
458 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
459 		return -ENOMEM;
460 	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
461 	if (rc < 0)
462 		goto out;
463 	rc = add_del_listener(info->snd_portid, mask, REGISTER);
464 out:
465 	free_cpumask_var(mask);
466 	return rc;
467 }
468 
469 static int cmd_attr_deregister_cpumask(struct genl_info *info)
470 {
471 	cpumask_var_t mask;
472 	int rc;
473 
474 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
475 		return -ENOMEM;
476 	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
477 	if (rc < 0)
478 		goto out;
479 	rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
480 out:
481 	free_cpumask_var(mask);
482 	return rc;
483 }
484 
485 static size_t taskstats_packet_size(void)
486 {
487 	size_t size;
488 
489 	size = nla_total_size(sizeof(u32)) +
490 		nla_total_size_64bit(sizeof(struct taskstats)) +
491 		nla_total_size(0);
492 
493 	return size;
494 }
495 
496 static int cmd_attr_pid(struct genl_info *info)
497 {
498 	struct taskstats *stats;
499 	struct sk_buff *rep_skb;
500 	size_t size;
501 	u32 pid;
502 	int rc;
503 
504 	size = taskstats_packet_size();
505 
506 	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
507 	if (rc < 0)
508 		return rc;
509 
510 	rc = -EINVAL;
511 	pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
512 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
513 	if (!stats)
514 		goto err;
515 
516 	rc = fill_stats_for_pid(pid, stats);
517 	if (rc < 0)
518 		goto err;
519 	return send_reply(rep_skb, info);
520 err:
521 	nlmsg_free(rep_skb);
522 	return rc;
523 }
524 
525 static int cmd_attr_tgid(struct genl_info *info)
526 {
527 	struct taskstats *stats;
528 	struct sk_buff *rep_skb;
529 	size_t size;
530 	u32 tgid;
531 	int rc;
532 
533 	size = taskstats_packet_size();
534 
535 	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
536 	if (rc < 0)
537 		return rc;
538 
539 	rc = -EINVAL;
540 	tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
541 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
542 	if (!stats)
543 		goto err;
544 
545 	rc = fill_stats_for_tgid(tgid, stats);
546 	if (rc < 0)
547 		goto err;
548 	return send_reply(rep_skb, info);
549 err:
550 	nlmsg_free(rep_skb);
551 	return rc;
552 }
553 
554 static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
555 {
556 	if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
557 		return cmd_attr_register_cpumask(info);
558 	else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
559 		return cmd_attr_deregister_cpumask(info);
560 	else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
561 		return cmd_attr_pid(info);
562 	else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
563 		return cmd_attr_tgid(info);
564 	else
565 		return -EINVAL;
566 }
567 
568 static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
569 {
570 	struct signal_struct *sig = tsk->signal;
571 	struct taskstats *stats;
572 
573 	if (sig->stats || thread_group_empty(tsk))
574 		goto ret;
575 
576 	/* No problem if kmem_cache_zalloc() fails */
577 	stats = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
578 
579 	spin_lock_irq(&tsk->sighand->siglock);
580 	if (!sig->stats) {
581 		sig->stats = stats;
582 		stats = NULL;
583 	}
584 	spin_unlock_irq(&tsk->sighand->siglock);
585 
586 	if (stats)
587 		kmem_cache_free(taskstats_cache, stats);
588 ret:
589 	return sig->stats;
590 }
591 
592 /* Send pid data out on exit */
593 void taskstats_exit(struct task_struct *tsk, int group_dead)
594 {
595 	int rc;
596 	struct listener_list *listeners;
597 	struct taskstats *stats;
598 	struct sk_buff *rep_skb;
599 	size_t size;
600 	int is_thread_group;
601 
602 	if (!family_registered)
603 		return;
604 
605 	/*
606 	 * Size includes space for nested attributes
607 	 */
608 	size = taskstats_packet_size();
609 
610 	is_thread_group = !!taskstats_tgid_alloc(tsk);
611 	if (is_thread_group) {
612 		/* PID + STATS + TGID + STATS */
613 		size = 2 * size;
614 		/* fill the tsk->signal->stats structure */
615 		fill_tgid_exit(tsk);
616 	}
617 
618 	listeners = raw_cpu_ptr(&listener_array);
619 	if (list_empty(&listeners->list))
620 		return;
621 
622 	rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
623 	if (rc < 0)
624 		return;
625 
626 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
627 			 task_pid_nr_ns(tsk, &init_pid_ns));
628 	if (!stats)
629 		goto err;
630 
631 	fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
632 
633 	/*
634 	 * Doesn't matter if tsk is the leader or the last group member leaving
635 	 */
636 	if (!is_thread_group || !group_dead)
637 		goto send;
638 
639 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
640 			 task_tgid_nr_ns(tsk, &init_pid_ns));
641 	if (!stats)
642 		goto err;
643 
644 	memcpy(stats, tsk->signal->stats, sizeof(*stats));
645 
646 send:
647 	send_cpu_listeners(rep_skb, listeners);
648 	return;
649 err:
650 	nlmsg_free(rep_skb);
651 }
652 
653 static const struct genl_ops taskstats_ops[] = {
654 	{
655 		.cmd		= TASKSTATS_CMD_GET,
656 		.doit		= taskstats_user_cmd,
657 		.policy		= taskstats_cmd_get_policy,
658 		.flags		= GENL_ADMIN_PERM,
659 	},
660 	{
661 		.cmd		= CGROUPSTATS_CMD_GET,
662 		.doit		= cgroupstats_user_cmd,
663 		.policy		= cgroupstats_cmd_get_policy,
664 	},
665 };
666 
667 static struct genl_family family __ro_after_init = {
668 	.name		= TASKSTATS_GENL_NAME,
669 	.version	= TASKSTATS_GENL_VERSION,
670 	.maxattr	= TASKSTATS_CMD_ATTR_MAX,
671 	.module		= THIS_MODULE,
672 	.ops		= taskstats_ops,
673 	.n_ops		= ARRAY_SIZE(taskstats_ops),
674 };
675 
676 /* Needed early in initialization */
677 void __init taskstats_init_early(void)
678 {
679 	unsigned int i;
680 
681 	taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
682 	for_each_possible_cpu(i) {
683 		INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
684 		init_rwsem(&(per_cpu(listener_array, i).sem));
685 	}
686 }
687 
688 static int __init taskstats_init(void)
689 {
690 	int rc;
691 
692 	rc = genl_register_family(&family);
693 	if (rc)
694 		return rc;
695 
696 	family_registered = 1;
697 	pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
698 	return 0;
699 }
700 
701 /*
702  * late initcall ensures initialization of statistics collection
703  * mechanisms precedes initialization of the taskstats interface
704  */
705 late_initcall(taskstats_init);
706