xref: /openbmc/linux/kernel/taskstats.c (revision 08b7cf13) 
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * taskstats.c - Export per-task statistics to userland
4  *
5  * Copyright (C) Shailabh Nagar, IBM Corp. 2006
6  *           (C) Balbir Singh,   IBM Corp. 2006
7  */
8 
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>
18 #include <linux/fs.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>
24 
25 /*
26  * Maximum length of a cpumask that can be specified in
27  * the TASKSTATS_CMD_ATTR_REGISTER/DEREGISTER_CPUMASK attribute
28  */
29 #define TASKSTATS_CPUMASK_MAXLEN	(100+6*NR_CPUS)
30 
31 static DEFINE_PER_CPU(__u32, taskstats_seqnum);
32 static int family_registered;
33 struct kmem_cache *taskstats_cache;
34 
35 static struct genl_family family;
36 
37 static const struct nla_policy taskstats_cmd_get_policy[] = {
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 },};
42 
43 static const struct nla_policy cgroupstats_cmd_get_policy[] = {
44 	[CGROUPSTATS_CMD_ATTR_FD] = { .type = NLA_U32 },
45 };
46 
47 struct listener {
48 	struct list_head list;
49 	pid_t pid;
50 	char valid;
51 };
52 
53 struct listener_list {
54 	struct rw_semaphore sem;
55 	struct list_head list;
56 };
57 static DEFINE_PER_CPU(struct listener_list, listener_array);
58 
59 enum actions {
60 	REGISTER,
61 	DEREGISTER,
62 	CPU_DONT_CARE
63 };
64 
65 static int prepare_reply(struct genl_info *info, u8 cmd, struct sk_buff **skbp,
66 				size_t size)
67 {
68 	struct sk_buff *skb;
69 	void *reply;
70 
71 	/*
72 	 * If new attributes are added, please revisit this allocation
73 	 */
74 	skb = genlmsg_new(size, GFP_KERNEL);
75 	if (!skb)
76 		return -ENOMEM;
77 
78 	if (!info) {
79 		int seq = this_cpu_inc_return(taskstats_seqnum) - 1;
80 
81 		reply = genlmsg_put(skb, 0, seq, &family, 0, cmd);
82 	} else
83 		reply = genlmsg_put_reply(skb, info, &family, 0, cmd);
84 	if (reply == NULL) {
85 		nlmsg_free(skb);
86 		return -EINVAL;
87 	}
88 
89 	*skbp = skb;
90 	return 0;
91 }
92 
93 /*
94  * Send taskstats data in @skb to listener with nl_pid @pid
95  */
96 static int send_reply(struct sk_buff *skb, struct genl_info *info)
97 {
98 	struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
99 	void *reply = genlmsg_data(genlhdr);
100 
101 	genlmsg_end(skb, reply);
102 
103 	return genlmsg_reply(skb, info);
104 }
105 
106 /*
107  * Send taskstats data in @skb to listeners registered for @cpu's exit data
108  */
109 static void send_cpu_listeners(struct sk_buff *skb,
110 					struct listener_list *listeners)
111 {
112 	struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
113 	struct listener *s, *tmp;
114 	struct sk_buff *skb_next, *skb_cur = skb;
115 	void *reply = genlmsg_data(genlhdr);
116 	int delcount = 0;
117 
118 	genlmsg_end(skb, reply);
119 
120 	down_read(&listeners->sem);
121 	list_for_each_entry(s, &listeners->list, list) {
122 		int rc;
123 
124 		skb_next = NULL;
125 		if (!list_is_last(&s->list, &listeners->list)) {
126 			skb_next = skb_clone(skb_cur, GFP_KERNEL);
127 			if (!skb_next)
128 				break;
129 		}
130 		rc = genlmsg_unicast(&init_net, skb_cur, s->pid);
131 		if (rc == -ECONNREFUSED) {
132 			s->valid = 0;
133 			delcount++;
134 		}
135 		skb_cur = skb_next;
136 	}
137 	up_read(&listeners->sem);
138 
139 	if (skb_cur)
140 		nlmsg_free(skb_cur);
141 
142 	if (!delcount)
143 		return;
144 
145 	/* Delete invalidated entries */
146 	down_write(&listeners->sem);
147 	list_for_each_entry_safe(s, tmp, &listeners->list, list) {
148 		if (!s->valid) {
149 			list_del(&s->list);
150 			kfree(s);
151 		}
152 	}
153 	up_write(&listeners->sem);
154 }
155 
156 static void fill_stats(struct user_namespace *user_ns,
157 		       struct pid_namespace *pid_ns,
158 		       struct task_struct *tsk, struct taskstats *stats)
159 {
160 	memset(stats, 0, sizeof(*stats));
161 	/*
162 	 * Each accounting subsystem adds calls to its functions to
163 	 * fill in relevant parts of struct taskstsats as follows
164 	 *
165 	 *	per-task-foo(stats, tsk);
166 	 */
167 
168 	delayacct_add_tsk(stats, tsk);
169 
170 	/* fill in basic acct fields */
171 	stats->version = TASKSTATS_VERSION;
172 	stats->nvcsw = tsk->nvcsw;
173 	stats->nivcsw = tsk->nivcsw;
174 	bacct_add_tsk(user_ns, pid_ns, stats, tsk);
175 
176 	/* fill in extended acct fields */
177 	xacct_add_tsk(stats, tsk);
178 }
179 
180 static int fill_stats_for_pid(pid_t pid, struct taskstats *stats)
181 {
182 	struct task_struct *tsk;
183 
184 	tsk = find_get_task_by_vpid(pid);
185 	if (!tsk)
186 		return -ESRCH;
187 	fill_stats(current_user_ns(), task_active_pid_ns(current), tsk, stats);
188 	put_task_struct(tsk);
189 	return 0;
190 }
191 
192 static int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
193 {
194 	struct task_struct *tsk, *first;
195 	unsigned long flags;
196 	int rc = -ESRCH;
197 	u64 delta, utime, stime;
198 	u64 start_time;
199 
200 	/*
201 	 * Add additional stats from live tasks except zombie thread group
202 	 * leaders who are already counted with the dead tasks
203 	 */
204 	rcu_read_lock();
205 	first = find_task_by_vpid(tgid);
206 
207 	if (!first || !lock_task_sighand(first, &flags))
208 		goto out;
209 
210 	if (first->signal->stats)
211 		memcpy(stats, first->signal->stats, sizeof(*stats));
212 	else
213 		memset(stats, 0, sizeof(*stats));
214 
215 	tsk = first;
216 	start_time = ktime_get_ns();
217 	do {
218 		if (tsk->exit_state)
219 			continue;
220 		/*
221 		 * Accounting subsystem can call its functions here to
222 		 * fill in relevant parts of struct taskstsats as follows
223 		 *
224 		 *	per-task-foo(stats, tsk);
225 		 */
226 		delayacct_add_tsk(stats, tsk);
227 
228 		/* calculate task elapsed time in nsec */
229 		delta = start_time - tsk->start_time;
230 		/* Convert to micro seconds */
231 		do_div(delta, NSEC_PER_USEC);
232 		stats->ac_etime += delta;
233 
234 		task_cputime(tsk, &utime, &stime);
235 		stats->ac_utime += div_u64(utime, NSEC_PER_USEC);
236 		stats->ac_stime += div_u64(stime, NSEC_PER_USEC);
237 
238 		stats->nvcsw += tsk->nvcsw;
239 		stats->nivcsw += tsk->nivcsw;
240 	} while_each_thread(first, tsk);
241 
242 	unlock_task_sighand(first, &flags);
243 	rc = 0;
244 out:
245 	rcu_read_unlock();
246 
247 	stats->version = TASKSTATS_VERSION;
248 	/*
249 	 * Accounting subsystems can also add calls here to modify
250 	 * fields of taskstats.
251 	 */
252 	return rc;
253 }
254 
255 static void fill_tgid_exit(struct task_struct *tsk)
256 {
257 	unsigned long flags;
258 
259 	spin_lock_irqsave(&tsk->sighand->siglock, flags);
260 	if (!tsk->signal->stats)
261 		goto ret;
262 
263 	/*
264 	 * Each accounting subsystem calls its functions here to
265 	 * accumalate its per-task stats for tsk, into the per-tgid structure
266 	 *
267 	 *	per-task-foo(tsk->signal->stats, tsk);
268 	 */
269 	delayacct_add_tsk(tsk->signal->stats, tsk);
270 ret:
271 	spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
272 	return;
273 }
274 
275 static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
276 {
277 	struct listener_list *listeners;
278 	struct listener *s, *tmp, *s2;
279 	unsigned int cpu;
280 	int ret = 0;
281 
282 	if (!cpumask_subset(mask, cpu_possible_mask))
283 		return -EINVAL;
284 
285 	if (current_user_ns() != &init_user_ns)
286 		return -EINVAL;
287 
288 	if (task_active_pid_ns(current) != &init_pid_ns)
289 		return -EINVAL;
290 
291 	if (isadd == REGISTER) {
292 		for_each_cpu(cpu, mask) {
293 			s = kmalloc_node(sizeof(struct listener),
294 					GFP_KERNEL, cpu_to_node(cpu));
295 			if (!s) {
296 				ret = -ENOMEM;
297 				goto cleanup;
298 			}
299 			s->pid = pid;
300 			s->valid = 1;
301 
302 			listeners = &per_cpu(listener_array, cpu);
303 			down_write(&listeners->sem);
304 			list_for_each_entry(s2, &listeners->list, list) {
305 				if (s2->pid == pid && s2->valid)
306 					goto exists;
307 			}
308 			list_add(&s->list, &listeners->list);
309 			s = NULL;
310 exists:
311 			up_write(&listeners->sem);
312 			kfree(s); /* nop if NULL */
313 		}
314 		return 0;
315 	}
316 
317 	/* Deregister or cleanup */
318 cleanup:
319 	for_each_cpu(cpu, mask) {
320 		listeners = &per_cpu(listener_array, cpu);
321 		down_write(&listeners->sem);
322 		list_for_each_entry_safe(s, tmp, &listeners->list, list) {
323 			if (s->pid == pid) {
324 				list_del(&s->list);
325 				kfree(s);
326 				break;
327 			}
328 		}
329 		up_write(&listeners->sem);
330 	}
331 	return ret;
332 }
333 
334 static int parse(struct nlattr *na, struct cpumask *mask)
335 {
336 	char *data;
337 	int len;
338 	int ret;
339 
340 	if (na == NULL)
341 		return 1;
342 	len = nla_len(na);
343 	if (len > TASKSTATS_CPUMASK_MAXLEN)
344 		return -E2BIG;
345 	if (len < 1)
346 		return -EINVAL;
347 	data = kmalloc(len, GFP_KERNEL);
348 	if (!data)
349 		return -ENOMEM;
350 	nla_strscpy(data, na, len);
351 	ret = cpulist_parse(data, mask);
352 	kfree(data);
353 	return ret;
354 }
355 
356 static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
357 {
358 	struct nlattr *na, *ret;
359 	int aggr;
360 
361 	aggr = (type == TASKSTATS_TYPE_PID)
362 			? TASKSTATS_TYPE_AGGR_PID
363 			: TASKSTATS_TYPE_AGGR_TGID;
364 
365 	na = nla_nest_start_noflag(skb, aggr);
366 	if (!na)
367 		goto err;
368 
369 	if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
370 		nla_nest_cancel(skb, na);
371 		goto err;
372 	}
373 	ret = nla_reserve_64bit(skb, TASKSTATS_TYPE_STATS,
374 				sizeof(struct taskstats), TASKSTATS_TYPE_NULL);
375 	if (!ret) {
376 		nla_nest_cancel(skb, na);
377 		goto err;
378 	}
379 	nla_nest_end(skb, na);
380 
381 	return nla_data(ret);
382 err:
383 	return NULL;
384 }
385 
386 static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
387 {
388 	int rc = 0;
389 	struct sk_buff *rep_skb;
390 	struct cgroupstats *stats;
391 	struct nlattr *na;
392 	size_t size;
393 	u32 fd;
394 	struct fd f;
395 
396 	na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
397 	if (!na)
398 		return -EINVAL;
399 
400 	fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
401 	f = fdget(fd);
402 	if (!f.file)
403 		return 0;
404 
405 	size = nla_total_size(sizeof(struct cgroupstats));
406 
407 	rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
408 				size);
409 	if (rc < 0)
410 		goto err;
411 
412 	na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
413 				sizeof(struct cgroupstats));
414 	if (na == NULL) {
415 		nlmsg_free(rep_skb);
416 		rc = -EMSGSIZE;
417 		goto err;
418 	}
419 
420 	stats = nla_data(na);
421 	memset(stats, 0, sizeof(*stats));
422 
423 	rc = cgroupstats_build(stats, f.file->f_path.dentry);
424 	if (rc < 0) {
425 		nlmsg_free(rep_skb);
426 		goto err;
427 	}
428 
429 	rc = send_reply(rep_skb, info);
430 
431 err:
432 	fdput(f);
433 	return rc;
434 }
435 
436 static int cmd_attr_register_cpumask(struct genl_info *info)
437 {
438 	cpumask_var_t mask;
439 	int rc;
440 
441 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
442 		return -ENOMEM;
443 	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
444 	if (rc < 0)
445 		goto out;
446 	rc = add_del_listener(info->snd_portid, mask, REGISTER);
447 out:
448 	free_cpumask_var(mask);
449 	return rc;
450 }
451 
452 static int cmd_attr_deregister_cpumask(struct genl_info *info)
453 {
454 	cpumask_var_t mask;
455 	int rc;
456 
457 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
458 		return -ENOMEM;
459 	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
460 	if (rc < 0)
461 		goto out;
462 	rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
463 out:
464 	free_cpumask_var(mask);
465 	return rc;
466 }
467 
468 static size_t taskstats_packet_size(void)
469 {
470 	size_t size;
471 
472 	size = nla_total_size(sizeof(u32)) +
473 		nla_total_size_64bit(sizeof(struct taskstats)) +
474 		nla_total_size(0);
475 
476 	return size;
477 }
478 
479 static int cmd_attr_pid(struct genl_info *info)
480 {
481 	struct taskstats *stats;
482 	struct sk_buff *rep_skb;
483 	size_t size;
484 	u32 pid;
485 	int rc;
486 
487 	size = taskstats_packet_size();
488 
489 	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
490 	if (rc < 0)
491 		return rc;
492 
493 	rc = -EINVAL;
494 	pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
495 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
496 	if (!stats)
497 		goto err;
498 
499 	rc = fill_stats_for_pid(pid, stats);
500 	if (rc < 0)
501 		goto err;
502 	return send_reply(rep_skb, info);
503 err:
504 	nlmsg_free(rep_skb);
505 	return rc;
506 }
507 
508 static int cmd_attr_tgid(struct genl_info *info)
509 {
510 	struct taskstats *stats;
511 	struct sk_buff *rep_skb;
512 	size_t size;
513 	u32 tgid;
514 	int rc;
515 
516 	size = taskstats_packet_size();
517 
518 	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
519 	if (rc < 0)
520 		return rc;
521 
522 	rc = -EINVAL;
523 	tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
524 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
525 	if (!stats)
526 		goto err;
527 
528 	rc = fill_stats_for_tgid(tgid, stats);
529 	if (rc < 0)
530 		goto err;
531 	return send_reply(rep_skb, info);
532 err:
533 	nlmsg_free(rep_skb);
534 	return rc;
535 }
536 
537 static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
538 {
539 	if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
540 		return cmd_attr_register_cpumask(info);
541 	else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
542 		return cmd_attr_deregister_cpumask(info);
543 	else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
544 		return cmd_attr_pid(info);
545 	else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
546 		return cmd_attr_tgid(info);
547 	else
548 		return -EINVAL;
549 }
550 
551 static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
552 {
553 	struct signal_struct *sig = tsk->signal;
554 	struct taskstats *stats_new, *stats;
555 
556 	/* Pairs with smp_store_release() below. */
557 	stats = smp_load_acquire(&sig->stats);
558 	if (stats || thread_group_empty(tsk))
559 		return stats;
560 
561 	/* No problem if kmem_cache_zalloc() fails */
562 	stats_new = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
563 
564 	spin_lock_irq(&tsk->sighand->siglock);
565 	stats = sig->stats;
566 	if (!stats) {
567 		/*
568 		 * Pairs with smp_store_release() above and order the
569 		 * kmem_cache_zalloc().
570 		 */
571 		smp_store_release(&sig->stats, stats_new);
572 		stats = stats_new;
573 		stats_new = NULL;
574 	}
575 	spin_unlock_irq(&tsk->sighand->siglock);
576 
577 	if (stats_new)
578 		kmem_cache_free(taskstats_cache, stats_new);
579 
580 	return stats;
581 }
582 
583 /* Send pid data out on exit */
584 void taskstats_exit(struct task_struct *tsk, int group_dead)
585 {
586 	int rc;
587 	struct listener_list *listeners;
588 	struct taskstats *stats;
589 	struct sk_buff *rep_skb;
590 	size_t size;
591 	int is_thread_group;
592 
593 	if (!family_registered)
594 		return;
595 
596 	/*
597 	 * Size includes space for nested attributes
598 	 */
599 	size = taskstats_packet_size();
600 
601 	is_thread_group = !!taskstats_tgid_alloc(tsk);
602 	if (is_thread_group) {
603 		/* PID + STATS + TGID + STATS */
604 		size = 2 * size;
605 		/* fill the tsk->signal->stats structure */
606 		fill_tgid_exit(tsk);
607 	}
608 
609 	listeners = raw_cpu_ptr(&listener_array);
610 	if (list_empty(&listeners->list))
611 		return;
612 
613 	rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
614 	if (rc < 0)
615 		return;
616 
617 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
618 			 task_pid_nr_ns(tsk, &init_pid_ns));
619 	if (!stats)
620 		goto err;
621 
622 	fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
623 
624 	/*
625 	 * Doesn't matter if tsk is the leader or the last group member leaving
626 	 */
627 	if (!is_thread_group || !group_dead)
628 		goto send;
629 
630 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
631 			 task_tgid_nr_ns(tsk, &init_pid_ns));
632 	if (!stats)
633 		goto err;
634 
635 	memcpy(stats, tsk->signal->stats, sizeof(*stats));
636 
637 send:
638 	send_cpu_listeners(rep_skb, listeners);
639 	return;
640 err:
641 	nlmsg_free(rep_skb);
642 }
643 
644 static const struct genl_ops taskstats_ops[] = {
645 	{
646 		.cmd		= TASKSTATS_CMD_GET,
647 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
648 		.doit		= taskstats_user_cmd,
649 		.policy		= taskstats_cmd_get_policy,
650 		.maxattr	= ARRAY_SIZE(taskstats_cmd_get_policy) - 1,
651 		.flags		= GENL_ADMIN_PERM,
652 	},
653 	{
654 		.cmd		= CGROUPSTATS_CMD_GET,
655 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
656 		.doit		= cgroupstats_user_cmd,
657 		.policy		= cgroupstats_cmd_get_policy,
658 		.maxattr	= ARRAY_SIZE(cgroupstats_cmd_get_policy) - 1,
659 	},
660 };
661 
662 static struct genl_family family __ro_after_init = {
663 	.name		= TASKSTATS_GENL_NAME,
664 	.version	= TASKSTATS_GENL_VERSION,
665 	.module		= THIS_MODULE,
666 	.ops		= taskstats_ops,
667 	.n_ops		= ARRAY_SIZE(taskstats_ops),
668 };
669 
670 /* Needed early in initialization */
671 void __init taskstats_init_early(void)
672 {
673 	unsigned int i;
674 
675 	taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
676 	for_each_possible_cpu(i) {
677 		INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
678 		init_rwsem(&(per_cpu(listener_array, i).sem));
679 	}
680 }
681 
682 static int __init taskstats_init(void)
683 {
684 	int rc;
685 
686 	rc = genl_register_family(&family);
687 	if (rc)
688 		return rc;
689 
690 	family_registered = 1;
691 	pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
692 	return 0;
693 }
694 
695 /*
696  * late initcall ensures initialization of statistics collection
697  * mechanisms precedes initialization of the taskstats interface
698  */
699 late_initcall(taskstats_init);
700