xref: /openbmc/linux/kernel/taskstats.c (revision bef7a78d) 
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 rc, delcount = 0;
117 
118 	genlmsg_end(skb, reply);
119 
120 	rc = 0;
121 	down_read(&listeners->sem);
122 	list_for_each_entry(s, &listeners->list, list) {
123 		skb_next = NULL;
124 		if (!list_is_last(&s->list, &listeners->list)) {
125 			skb_next = skb_clone(skb_cur, GFP_KERNEL);
126 			if (!skb_next)
127 				break;
128 		}
129 		rc = genlmsg_unicast(&init_net, skb_cur, s->pid);
130 		if (rc == -ECONNREFUSED) {
131 			s->valid = 0;
132 			delcount++;
133 		}
134 		skb_cur = skb_next;
135 	}
136 	up_read(&listeners->sem);
137 
138 	if (skb_cur)
139 		nlmsg_free(skb_cur);
140 
141 	if (!delcount)
142 		return;
143 
144 	/* Delete invalidated entries */
145 	down_write(&listeners->sem);
146 	list_for_each_entry_safe(s, tmp, &listeners->list, list) {
147 		if (!s->valid) {
148 			list_del(&s->list);
149 			kfree(s);
150 		}
151 	}
152 	up_write(&listeners->sem);
153 }
154 
155 static void fill_stats(struct user_namespace *user_ns,
156 		       struct pid_namespace *pid_ns,
157 		       struct task_struct *tsk, struct taskstats *stats)
158 {
159 	memset(stats, 0, sizeof(*stats));
160 	/*
161 	 * Each accounting subsystem adds calls to its functions to
162 	 * fill in relevant parts of struct taskstsats as follows
163 	 *
164 	 *	per-task-foo(stats, tsk);
165 	 */
166 
167 	delayacct_add_tsk(stats, tsk);
168 
169 	/* fill in basic acct fields */
170 	stats->version = TASKSTATS_VERSION;
171 	stats->nvcsw = tsk->nvcsw;
172 	stats->nivcsw = tsk->nivcsw;
173 	bacct_add_tsk(user_ns, pid_ns, stats, tsk);
174 
175 	/* fill in extended acct fields */
176 	xacct_add_tsk(stats, tsk);
177 }
178 
179 static int fill_stats_for_pid(pid_t pid, struct taskstats *stats)
180 {
181 	struct task_struct *tsk;
182 
183 	tsk = find_get_task_by_vpid(pid);
184 	if (!tsk)
185 		return -ESRCH;
186 	fill_stats(current_user_ns(), task_active_pid_ns(current), tsk, stats);
187 	put_task_struct(tsk);
188 	return 0;
189 }
190 
191 static int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
192 {
193 	struct task_struct *tsk, *first;
194 	unsigned long flags;
195 	int rc = -ESRCH;
196 	u64 delta, utime, stime;
197 	u64 start_time;
198 
199 	/*
200 	 * Add additional stats from live tasks except zombie thread group
201 	 * leaders who are already counted with the dead tasks
202 	 */
203 	rcu_read_lock();
204 	first = find_task_by_vpid(tgid);
205 
206 	if (!first || !lock_task_sighand(first, &flags))
207 		goto out;
208 
209 	if (first->signal->stats)
210 		memcpy(stats, first->signal->stats, sizeof(*stats));
211 	else
212 		memset(stats, 0, sizeof(*stats));
213 
214 	tsk = first;
215 	start_time = ktime_get_ns();
216 	do {
217 		if (tsk->exit_state)
218 			continue;
219 		/*
220 		 * Accounting subsystem can call its functions here to
221 		 * fill in relevant parts of struct taskstsats as follows
222 		 *
223 		 *	per-task-foo(stats, tsk);
224 		 */
225 		delayacct_add_tsk(stats, tsk);
226 
227 		/* calculate task elapsed time in nsec */
228 		delta = start_time - tsk->start_time;
229 		/* Convert to micro seconds */
230 		do_div(delta, NSEC_PER_USEC);
231 		stats->ac_etime += delta;
232 
233 		task_cputime(tsk, &utime, &stime);
234 		stats->ac_utime += div_u64(utime, NSEC_PER_USEC);
235 		stats->ac_stime += div_u64(stime, NSEC_PER_USEC);
236 
237 		stats->nvcsw += tsk->nvcsw;
238 		stats->nivcsw += tsk->nivcsw;
239 	} while_each_thread(first, tsk);
240 
241 	unlock_task_sighand(first, &flags);
242 	rc = 0;
243 out:
244 	rcu_read_unlock();
245 
246 	stats->version = TASKSTATS_VERSION;
247 	/*
248 	 * Accounting subsystems can also add calls here to modify
249 	 * fields of taskstats.
250 	 */
251 	return rc;
252 }
253 
254 static void fill_tgid_exit(struct task_struct *tsk)
255 {
256 	unsigned long flags;
257 
258 	spin_lock_irqsave(&tsk->sighand->siglock, flags);
259 	if (!tsk->signal->stats)
260 		goto ret;
261 
262 	/*
263 	 * Each accounting subsystem calls its functions here to
264 	 * accumalate its per-task stats for tsk, into the per-tgid structure
265 	 *
266 	 *	per-task-foo(tsk->signal->stats, tsk);
267 	 */
268 	delayacct_add_tsk(tsk->signal->stats, tsk);
269 ret:
270 	spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
271 	return;
272 }
273 
274 static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
275 {
276 	struct listener_list *listeners;
277 	struct listener *s, *tmp, *s2;
278 	unsigned int cpu;
279 	int ret = 0;
280 
281 	if (!cpumask_subset(mask, cpu_possible_mask))
282 		return -EINVAL;
283 
284 	if (current_user_ns() != &init_user_ns)
285 		return -EINVAL;
286 
287 	if (task_active_pid_ns(current) != &init_pid_ns)
288 		return -EINVAL;
289 
290 	if (isadd == REGISTER) {
291 		for_each_cpu(cpu, mask) {
292 			s = kmalloc_node(sizeof(struct listener),
293 					GFP_KERNEL, cpu_to_node(cpu));
294 			if (!s) {
295 				ret = -ENOMEM;
296 				goto cleanup;
297 			}
298 			s->pid = pid;
299 			s->valid = 1;
300 
301 			listeners = &per_cpu(listener_array, cpu);
302 			down_write(&listeners->sem);
303 			list_for_each_entry(s2, &listeners->list, list) {
304 				if (s2->pid == pid && s2->valid)
305 					goto exists;
306 			}
307 			list_add(&s->list, &listeners->list);
308 			s = NULL;
309 exists:
310 			up_write(&listeners->sem);
311 			kfree(s); /* nop if NULL */
312 		}
313 		return 0;
314 	}
315 
316 	/* Deregister or cleanup */
317 cleanup:
318 	for_each_cpu(cpu, mask) {
319 		listeners = &per_cpu(listener_array, cpu);
320 		down_write(&listeners->sem);
321 		list_for_each_entry_safe(s, tmp, &listeners->list, list) {
322 			if (s->pid == pid) {
323 				list_del(&s->list);
324 				kfree(s);
325 				break;
326 			}
327 		}
328 		up_write(&listeners->sem);
329 	}
330 	return ret;
331 }
332 
333 static int parse(struct nlattr *na, struct cpumask *mask)
334 {
335 	char *data;
336 	int len;
337 	int ret;
338 
339 	if (na == NULL)
340 		return 1;
341 	len = nla_len(na);
342 	if (len > TASKSTATS_CPUMASK_MAXLEN)
343 		return -E2BIG;
344 	if (len < 1)
345 		return -EINVAL;
346 	data = kmalloc(len, GFP_KERNEL);
347 	if (!data)
348 		return -ENOMEM;
349 	nla_strscpy(data, na, len);
350 	ret = cpulist_parse(data, mask);
351 	kfree(data);
352 	return ret;
353 }
354 
355 static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
356 {
357 	struct nlattr *na, *ret;
358 	int aggr;
359 
360 	aggr = (type == TASKSTATS_TYPE_PID)
361 			? TASKSTATS_TYPE_AGGR_PID
362 			: TASKSTATS_TYPE_AGGR_TGID;
363 
364 	na = nla_nest_start_noflag(skb, aggr);
365 	if (!na)
366 		goto err;
367 
368 	if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
369 		nla_nest_cancel(skb, na);
370 		goto err;
371 	}
372 	ret = nla_reserve_64bit(skb, TASKSTATS_TYPE_STATS,
373 				sizeof(struct taskstats), TASKSTATS_TYPE_NULL);
374 	if (!ret) {
375 		nla_nest_cancel(skb, na);
376 		goto err;
377 	}
378 	nla_nest_end(skb, na);
379 
380 	return nla_data(ret);
381 err:
382 	return NULL;
383 }
384 
385 static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
386 {
387 	int rc = 0;
388 	struct sk_buff *rep_skb;
389 	struct cgroupstats *stats;
390 	struct nlattr *na;
391 	size_t size;
392 	u32 fd;
393 	struct fd f;
394 
395 	na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
396 	if (!na)
397 		return -EINVAL;
398 
399 	fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
400 	f = fdget(fd);
401 	if (!f.file)
402 		return 0;
403 
404 	size = nla_total_size(sizeof(struct cgroupstats));
405 
406 	rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
407 				size);
408 	if (rc < 0)
409 		goto err;
410 
411 	na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
412 				sizeof(struct cgroupstats));
413 	if (na == NULL) {
414 		nlmsg_free(rep_skb);
415 		rc = -EMSGSIZE;
416 		goto err;
417 	}
418 
419 	stats = nla_data(na);
420 	memset(stats, 0, sizeof(*stats));
421 
422 	rc = cgroupstats_build(stats, f.file->f_path.dentry);
423 	if (rc < 0) {
424 		nlmsg_free(rep_skb);
425 		goto err;
426 	}
427 
428 	rc = send_reply(rep_skb, info);
429 
430 err:
431 	fdput(f);
432 	return rc;
433 }
434 
435 static int cmd_attr_register_cpumask(struct genl_info *info)
436 {
437 	cpumask_var_t mask;
438 	int rc;
439 
440 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
441 		return -ENOMEM;
442 	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
443 	if (rc < 0)
444 		goto out;
445 	rc = add_del_listener(info->snd_portid, mask, REGISTER);
446 out:
447 	free_cpumask_var(mask);
448 	return rc;
449 }
450 
451 static int cmd_attr_deregister_cpumask(struct genl_info *info)
452 {
453 	cpumask_var_t mask;
454 	int rc;
455 
456 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
457 		return -ENOMEM;
458 	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
459 	if (rc < 0)
460 		goto out;
461 	rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
462 out:
463 	free_cpumask_var(mask);
464 	return rc;
465 }
466 
467 static size_t taskstats_packet_size(void)
468 {
469 	size_t size;
470 
471 	size = nla_total_size(sizeof(u32)) +
472 		nla_total_size_64bit(sizeof(struct taskstats)) +
473 		nla_total_size(0);
474 
475 	return size;
476 }
477 
478 static int cmd_attr_pid(struct genl_info *info)
479 {
480 	struct taskstats *stats;
481 	struct sk_buff *rep_skb;
482 	size_t size;
483 	u32 pid;
484 	int rc;
485 
486 	size = taskstats_packet_size();
487 
488 	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
489 	if (rc < 0)
490 		return rc;
491 
492 	rc = -EINVAL;
493 	pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
494 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
495 	if (!stats)
496 		goto err;
497 
498 	rc = fill_stats_for_pid(pid, stats);
499 	if (rc < 0)
500 		goto err;
501 	return send_reply(rep_skb, info);
502 err:
503 	nlmsg_free(rep_skb);
504 	return rc;
505 }
506 
507 static int cmd_attr_tgid(struct genl_info *info)
508 {
509 	struct taskstats *stats;
510 	struct sk_buff *rep_skb;
511 	size_t size;
512 	u32 tgid;
513 	int rc;
514 
515 	size = taskstats_packet_size();
516 
517 	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
518 	if (rc < 0)
519 		return rc;
520 
521 	rc = -EINVAL;
522 	tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
523 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
524 	if (!stats)
525 		goto err;
526 
527 	rc = fill_stats_for_tgid(tgid, stats);
528 	if (rc < 0)
529 		goto err;
530 	return send_reply(rep_skb, info);
531 err:
532 	nlmsg_free(rep_skb);
533 	return rc;
534 }
535 
536 static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
537 {
538 	if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
539 		return cmd_attr_register_cpumask(info);
540 	else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
541 		return cmd_attr_deregister_cpumask(info);
542 	else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
543 		return cmd_attr_pid(info);
544 	else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
545 		return cmd_attr_tgid(info);
546 	else
547 		return -EINVAL;
548 }
549 
550 static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
551 {
552 	struct signal_struct *sig = tsk->signal;
553 	struct taskstats *stats_new, *stats;
554 
555 	/* Pairs with smp_store_release() below. */
556 	stats = smp_load_acquire(&sig->stats);
557 	if (stats || thread_group_empty(tsk))
558 		return stats;
559 
560 	/* No problem if kmem_cache_zalloc() fails */
561 	stats_new = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
562 
563 	spin_lock_irq(&tsk->sighand->siglock);
564 	stats = sig->stats;
565 	if (!stats) {
566 		/*
567 		 * Pairs with smp_store_release() above and order the
568 		 * kmem_cache_zalloc().
569 		 */
570 		smp_store_release(&sig->stats, stats_new);
571 		stats = stats_new;
572 		stats_new = NULL;
573 	}
574 	spin_unlock_irq(&tsk->sighand->siglock);
575 
576 	if (stats_new)
577 		kmem_cache_free(taskstats_cache, stats_new);
578 
579 	return stats;
580 }
581 
582 /* Send pid data out on exit */
583 void taskstats_exit(struct task_struct *tsk, int group_dead)
584 {
585 	int rc;
586 	struct listener_list *listeners;
587 	struct taskstats *stats;
588 	struct sk_buff *rep_skb;
589 	size_t size;
590 	int is_thread_group;
591 
592 	if (!family_registered)
593 		return;
594 
595 	/*
596 	 * Size includes space for nested attributes
597 	 */
598 	size = taskstats_packet_size();
599 
600 	is_thread_group = !!taskstats_tgid_alloc(tsk);
601 	if (is_thread_group) {
602 		/* PID + STATS + TGID + STATS */
603 		size = 2 * size;
604 		/* fill the tsk->signal->stats structure */
605 		fill_tgid_exit(tsk);
606 	}
607 
608 	listeners = raw_cpu_ptr(&listener_array);
609 	if (list_empty(&listeners->list))
610 		return;
611 
612 	rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
613 	if (rc < 0)
614 		return;
615 
616 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
617 			 task_pid_nr_ns(tsk, &init_pid_ns));
618 	if (!stats)
619 		goto err;
620 
621 	fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
622 
623 	/*
624 	 * Doesn't matter if tsk is the leader or the last group member leaving
625 	 */
626 	if (!is_thread_group || !group_dead)
627 		goto send;
628 
629 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
630 			 task_tgid_nr_ns(tsk, &init_pid_ns));
631 	if (!stats)
632 		goto err;
633 
634 	memcpy(stats, tsk->signal->stats, sizeof(*stats));
635 
636 send:
637 	send_cpu_listeners(rep_skb, listeners);
638 	return;
639 err:
640 	nlmsg_free(rep_skb);
641 }
642 
643 static const struct genl_ops taskstats_ops[] = {
644 	{
645 		.cmd		= TASKSTATS_CMD_GET,
646 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
647 		.doit		= taskstats_user_cmd,
648 		.policy		= taskstats_cmd_get_policy,
649 		.maxattr	= ARRAY_SIZE(taskstats_cmd_get_policy) - 1,
650 		.flags		= GENL_ADMIN_PERM,
651 	},
652 	{
653 		.cmd		= CGROUPSTATS_CMD_GET,
654 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
655 		.doit		= cgroupstats_user_cmd,
656 		.policy		= cgroupstats_cmd_get_policy,
657 		.maxattr	= ARRAY_SIZE(cgroupstats_cmd_get_policy) - 1,
658 	},
659 };
660 
661 static struct genl_family family __ro_after_init = {
662 	.name		= TASKSTATS_GENL_NAME,
663 	.version	= TASKSTATS_GENL_VERSION,
664 	.module		= THIS_MODULE,
665 	.ops		= taskstats_ops,
666 	.n_ops		= ARRAY_SIZE(taskstats_ops),
667 };
668 
669 /* Needed early in initialization */
670 void __init taskstats_init_early(void)
671 {
672 	unsigned int i;
673 
674 	taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
675 	for_each_possible_cpu(i) {
676 		INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
677 		init_rwsem(&(per_cpu(listener_array, i).sem));
678 	}
679 }
680 
681 static int __init taskstats_init(void)
682 {
683 	int rc;
684 
685 	rc = genl_register_family(&family);
686 	if (rc)
687 		return rc;
688 
689 	family_registered = 1;
690 	pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
691 	return 0;
692 }
693 
694 /*
695  * late initcall ensures initialization of statistics collection
696  * mechanisms precedes initialization of the taskstats interface
697  */
698 late_initcall(taskstats_init);
699