xref: /openbmc/linux/kernel/taskstats.c (revision be709d48)
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 	tsk = find_get_task_by_vpid(pid);
198 	if (!tsk)
199 		return -ESRCH;
200 	fill_stats(current_user_ns(), task_active_pid_ns(current), tsk, stats);
201 	put_task_struct(tsk);
202 	return 0;
203 }
204 
205 static int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
206 {
207 	struct task_struct *tsk, *first;
208 	unsigned long flags;
209 	int rc = -ESRCH;
210 	u64 delta, utime, stime;
211 	u64 start_time;
212 
213 	/*
214 	 * Add additional stats from live tasks except zombie thread group
215 	 * leaders who are already counted with the dead tasks
216 	 */
217 	rcu_read_lock();
218 	first = find_task_by_vpid(tgid);
219 
220 	if (!first || !lock_task_sighand(first, &flags))
221 		goto out;
222 
223 	if (first->signal->stats)
224 		memcpy(stats, first->signal->stats, sizeof(*stats));
225 	else
226 		memset(stats, 0, sizeof(*stats));
227 
228 	tsk = first;
229 	start_time = ktime_get_ns();
230 	do {
231 		if (tsk->exit_state)
232 			continue;
233 		/*
234 		 * Accounting subsystem can call its functions here to
235 		 * fill in relevant parts of struct taskstsats as follows
236 		 *
237 		 *	per-task-foo(stats, tsk);
238 		 */
239 		delayacct_add_tsk(stats, tsk);
240 
241 		/* calculate task elapsed time in nsec */
242 		delta = start_time - tsk->start_time;
243 		/* Convert to micro seconds */
244 		do_div(delta, NSEC_PER_USEC);
245 		stats->ac_etime += delta;
246 
247 		task_cputime(tsk, &utime, &stime);
248 		stats->ac_utime += div_u64(utime, NSEC_PER_USEC);
249 		stats->ac_stime += div_u64(stime, NSEC_PER_USEC);
250 
251 		stats->nvcsw += tsk->nvcsw;
252 		stats->nivcsw += tsk->nivcsw;
253 	} while_each_thread(first, tsk);
254 
255 	unlock_task_sighand(first, &flags);
256 	rc = 0;
257 out:
258 	rcu_read_unlock();
259 
260 	stats->version = TASKSTATS_VERSION;
261 	/*
262 	 * Accounting subsystems can also add calls here to modify
263 	 * fields of taskstats.
264 	 */
265 	return rc;
266 }
267 
268 static void fill_tgid_exit(struct task_struct *tsk)
269 {
270 	unsigned long flags;
271 
272 	spin_lock_irqsave(&tsk->sighand->siglock, flags);
273 	if (!tsk->signal->stats)
274 		goto ret;
275 
276 	/*
277 	 * Each accounting subsystem calls its functions here to
278 	 * accumalate its per-task stats for tsk, into the per-tgid structure
279 	 *
280 	 *	per-task-foo(tsk->signal->stats, tsk);
281 	 */
282 	delayacct_add_tsk(tsk->signal->stats, tsk);
283 ret:
284 	spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
285 	return;
286 }
287 
288 static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
289 {
290 	struct listener_list *listeners;
291 	struct listener *s, *tmp, *s2;
292 	unsigned int cpu;
293 	int ret = 0;
294 
295 	if (!cpumask_subset(mask, cpu_possible_mask))
296 		return -EINVAL;
297 
298 	if (current_user_ns() != &init_user_ns)
299 		return -EINVAL;
300 
301 	if (task_active_pid_ns(current) != &init_pid_ns)
302 		return -EINVAL;
303 
304 	if (isadd == REGISTER) {
305 		for_each_cpu(cpu, mask) {
306 			s = kmalloc_node(sizeof(struct listener),
307 					GFP_KERNEL, cpu_to_node(cpu));
308 			if (!s) {
309 				ret = -ENOMEM;
310 				goto cleanup;
311 			}
312 			s->pid = pid;
313 			s->valid = 1;
314 
315 			listeners = &per_cpu(listener_array, cpu);
316 			down_write(&listeners->sem);
317 			list_for_each_entry(s2, &listeners->list, list) {
318 				if (s2->pid == pid && s2->valid)
319 					goto exists;
320 			}
321 			list_add(&s->list, &listeners->list);
322 			s = NULL;
323 exists:
324 			up_write(&listeners->sem);
325 			kfree(s); /* nop if NULL */
326 		}
327 		return 0;
328 	}
329 
330 	/* Deregister or cleanup */
331 cleanup:
332 	for_each_cpu(cpu, mask) {
333 		listeners = &per_cpu(listener_array, cpu);
334 		down_write(&listeners->sem);
335 		list_for_each_entry_safe(s, tmp, &listeners->list, list) {
336 			if (s->pid == pid) {
337 				list_del(&s->list);
338 				kfree(s);
339 				break;
340 			}
341 		}
342 		up_write(&listeners->sem);
343 	}
344 	return ret;
345 }
346 
347 static int parse(struct nlattr *na, struct cpumask *mask)
348 {
349 	char *data;
350 	int len;
351 	int ret;
352 
353 	if (na == NULL)
354 		return 1;
355 	len = nla_len(na);
356 	if (len > TASKSTATS_CPUMASK_MAXLEN)
357 		return -E2BIG;
358 	if (len < 1)
359 		return -EINVAL;
360 	data = kmalloc(len, GFP_KERNEL);
361 	if (!data)
362 		return -ENOMEM;
363 	nla_strlcpy(data, na, len);
364 	ret = cpulist_parse(data, mask);
365 	kfree(data);
366 	return ret;
367 }
368 
369 static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
370 {
371 	struct nlattr *na, *ret;
372 	int aggr;
373 
374 	aggr = (type == TASKSTATS_TYPE_PID)
375 			? TASKSTATS_TYPE_AGGR_PID
376 			: TASKSTATS_TYPE_AGGR_TGID;
377 
378 	na = nla_nest_start(skb, aggr);
379 	if (!na)
380 		goto err;
381 
382 	if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
383 		nla_nest_cancel(skb, na);
384 		goto err;
385 	}
386 	ret = nla_reserve_64bit(skb, TASKSTATS_TYPE_STATS,
387 				sizeof(struct taskstats), TASKSTATS_TYPE_NULL);
388 	if (!ret) {
389 		nla_nest_cancel(skb, na);
390 		goto err;
391 	}
392 	nla_nest_end(skb, na);
393 
394 	return nla_data(ret);
395 err:
396 	return NULL;
397 }
398 
399 static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
400 {
401 	int rc = 0;
402 	struct sk_buff *rep_skb;
403 	struct cgroupstats *stats;
404 	struct nlattr *na;
405 	size_t size;
406 	u32 fd;
407 	struct fd f;
408 
409 	na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
410 	if (!na)
411 		return -EINVAL;
412 
413 	fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
414 	f = fdget(fd);
415 	if (!f.file)
416 		return 0;
417 
418 	size = nla_total_size(sizeof(struct cgroupstats));
419 
420 	rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
421 				size);
422 	if (rc < 0)
423 		goto err;
424 
425 	na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
426 				sizeof(struct cgroupstats));
427 	if (na == NULL) {
428 		nlmsg_free(rep_skb);
429 		rc = -EMSGSIZE;
430 		goto err;
431 	}
432 
433 	stats = nla_data(na);
434 	memset(stats, 0, sizeof(*stats));
435 
436 	rc = cgroupstats_build(stats, f.file->f_path.dentry);
437 	if (rc < 0) {
438 		nlmsg_free(rep_skb);
439 		goto err;
440 	}
441 
442 	rc = send_reply(rep_skb, info);
443 
444 err:
445 	fdput(f);
446 	return rc;
447 }
448 
449 static int cmd_attr_register_cpumask(struct genl_info *info)
450 {
451 	cpumask_var_t mask;
452 	int rc;
453 
454 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
455 		return -ENOMEM;
456 	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
457 	if (rc < 0)
458 		goto out;
459 	rc = add_del_listener(info->snd_portid, mask, REGISTER);
460 out:
461 	free_cpumask_var(mask);
462 	return rc;
463 }
464 
465 static int cmd_attr_deregister_cpumask(struct genl_info *info)
466 {
467 	cpumask_var_t mask;
468 	int rc;
469 
470 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
471 		return -ENOMEM;
472 	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
473 	if (rc < 0)
474 		goto out;
475 	rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
476 out:
477 	free_cpumask_var(mask);
478 	return rc;
479 }
480 
481 static size_t taskstats_packet_size(void)
482 {
483 	size_t size;
484 
485 	size = nla_total_size(sizeof(u32)) +
486 		nla_total_size_64bit(sizeof(struct taskstats)) +
487 		nla_total_size(0);
488 
489 	return size;
490 }
491 
492 static int cmd_attr_pid(struct genl_info *info)
493 {
494 	struct taskstats *stats;
495 	struct sk_buff *rep_skb;
496 	size_t size;
497 	u32 pid;
498 	int rc;
499 
500 	size = taskstats_packet_size();
501 
502 	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
503 	if (rc < 0)
504 		return rc;
505 
506 	rc = -EINVAL;
507 	pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
508 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
509 	if (!stats)
510 		goto err;
511 
512 	rc = fill_stats_for_pid(pid, stats);
513 	if (rc < 0)
514 		goto err;
515 	return send_reply(rep_skb, info);
516 err:
517 	nlmsg_free(rep_skb);
518 	return rc;
519 }
520 
521 static int cmd_attr_tgid(struct genl_info *info)
522 {
523 	struct taskstats *stats;
524 	struct sk_buff *rep_skb;
525 	size_t size;
526 	u32 tgid;
527 	int rc;
528 
529 	size = taskstats_packet_size();
530 
531 	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
532 	if (rc < 0)
533 		return rc;
534 
535 	rc = -EINVAL;
536 	tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
537 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
538 	if (!stats)
539 		goto err;
540 
541 	rc = fill_stats_for_tgid(tgid, stats);
542 	if (rc < 0)
543 		goto err;
544 	return send_reply(rep_skb, info);
545 err:
546 	nlmsg_free(rep_skb);
547 	return rc;
548 }
549 
550 static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
551 {
552 	if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
553 		return cmd_attr_register_cpumask(info);
554 	else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
555 		return cmd_attr_deregister_cpumask(info);
556 	else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
557 		return cmd_attr_pid(info);
558 	else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
559 		return cmd_attr_tgid(info);
560 	else
561 		return -EINVAL;
562 }
563 
564 static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
565 {
566 	struct signal_struct *sig = tsk->signal;
567 	struct taskstats *stats;
568 
569 	if (sig->stats || thread_group_empty(tsk))
570 		goto ret;
571 
572 	/* No problem if kmem_cache_zalloc() fails */
573 	stats = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
574 
575 	spin_lock_irq(&tsk->sighand->siglock);
576 	if (!sig->stats) {
577 		sig->stats = stats;
578 		stats = NULL;
579 	}
580 	spin_unlock_irq(&tsk->sighand->siglock);
581 
582 	if (stats)
583 		kmem_cache_free(taskstats_cache, stats);
584 ret:
585 	return sig->stats;
586 }
587 
588 /* Send pid data out on exit */
589 void taskstats_exit(struct task_struct *tsk, int group_dead)
590 {
591 	int rc;
592 	struct listener_list *listeners;
593 	struct taskstats *stats;
594 	struct sk_buff *rep_skb;
595 	size_t size;
596 	int is_thread_group;
597 
598 	if (!family_registered)
599 		return;
600 
601 	/*
602 	 * Size includes space for nested attributes
603 	 */
604 	size = taskstats_packet_size();
605 
606 	is_thread_group = !!taskstats_tgid_alloc(tsk);
607 	if (is_thread_group) {
608 		/* PID + STATS + TGID + STATS */
609 		size = 2 * size;
610 		/* fill the tsk->signal->stats structure */
611 		fill_tgid_exit(tsk);
612 	}
613 
614 	listeners = raw_cpu_ptr(&listener_array);
615 	if (list_empty(&listeners->list))
616 		return;
617 
618 	rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
619 	if (rc < 0)
620 		return;
621 
622 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
623 			 task_pid_nr_ns(tsk, &init_pid_ns));
624 	if (!stats)
625 		goto err;
626 
627 	fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
628 
629 	/*
630 	 * Doesn't matter if tsk is the leader or the last group member leaving
631 	 */
632 	if (!is_thread_group || !group_dead)
633 		goto send;
634 
635 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
636 			 task_tgid_nr_ns(tsk, &init_pid_ns));
637 	if (!stats)
638 		goto err;
639 
640 	memcpy(stats, tsk->signal->stats, sizeof(*stats));
641 
642 send:
643 	send_cpu_listeners(rep_skb, listeners);
644 	return;
645 err:
646 	nlmsg_free(rep_skb);
647 }
648 
649 static const struct genl_ops taskstats_ops[] = {
650 	{
651 		.cmd		= TASKSTATS_CMD_GET,
652 		.doit		= taskstats_user_cmd,
653 		.policy		= taskstats_cmd_get_policy,
654 		.flags		= GENL_ADMIN_PERM,
655 	},
656 	{
657 		.cmd		= CGROUPSTATS_CMD_GET,
658 		.doit		= cgroupstats_user_cmd,
659 		.policy		= cgroupstats_cmd_get_policy,
660 	},
661 };
662 
663 static struct genl_family family __ro_after_init = {
664 	.name		= TASKSTATS_GENL_NAME,
665 	.version	= TASKSTATS_GENL_VERSION,
666 	.maxattr	= TASKSTATS_CMD_ATTR_MAX,
667 	.module		= THIS_MODULE,
668 	.ops		= taskstats_ops,
669 	.n_ops		= ARRAY_SIZE(taskstats_ops),
670 };
671 
672 /* Needed early in initialization */
673 void __init taskstats_init_early(void)
674 {
675 	unsigned int i;
676 
677 	taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
678 	for_each_possible_cpu(i) {
679 		INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
680 		init_rwsem(&(per_cpu(listener_array, i).sem));
681 	}
682 }
683 
684 static int __init taskstats_init(void)
685 {
686 	int rc;
687 
688 	rc = genl_register_family(&family);
689 	if (rc)
690 		return rc;
691 
692 	family_registered = 1;
693 	pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
694 	return 0;
695 }
696 
697 /*
698  * late initcall ensures initialization of statistics collection
699  * mechanisms precedes initialization of the taskstats interface
700  */
701 late_initcall(taskstats_init);
702