xref: /openbmc/linux/kernel/taskstats.c (revision 565d76cb)
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 <net/genetlink.h>
31 #include <asm/atomic.h>
32 
33 /*
34  * Maximum length of a cpumask that can be specified in
35  * the TASKSTATS_CMD_ATTR_REGISTER/DEREGISTER_CPUMASK attribute
36  */
37 #define TASKSTATS_CPUMASK_MAXLEN	(100+6*NR_CPUS)
38 
39 static DEFINE_PER_CPU(__u32, taskstats_seqnum);
40 static int family_registered;
41 struct kmem_cache *taskstats_cache;
42 
43 static struct genl_family family = {
44 	.id		= GENL_ID_GENERATE,
45 	.name		= TASKSTATS_GENL_NAME,
46 	.version	= TASKSTATS_GENL_VERSION,
47 	.maxattr	= TASKSTATS_CMD_ATTR_MAX,
48 };
49 
50 static const struct nla_policy taskstats_cmd_get_policy[TASKSTATS_CMD_ATTR_MAX+1] = {
51 	[TASKSTATS_CMD_ATTR_PID]  = { .type = NLA_U32 },
52 	[TASKSTATS_CMD_ATTR_TGID] = { .type = NLA_U32 },
53 	[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK] = { .type = NLA_STRING },
54 	[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK] = { .type = NLA_STRING },};
55 
56 static const struct nla_policy cgroupstats_cmd_get_policy[CGROUPSTATS_CMD_ATTR_MAX+1] = {
57 	[CGROUPSTATS_CMD_ATTR_FD] = { .type = NLA_U32 },
58 };
59 
60 struct listener {
61 	struct list_head list;
62 	pid_t pid;
63 	char valid;
64 };
65 
66 struct listener_list {
67 	struct rw_semaphore sem;
68 	struct list_head list;
69 };
70 static DEFINE_PER_CPU(struct listener_list, listener_array);
71 
72 enum actions {
73 	REGISTER,
74 	DEREGISTER,
75 	CPU_DONT_CARE
76 };
77 
78 static int prepare_reply(struct genl_info *info, u8 cmd, struct sk_buff **skbp,
79 				size_t size)
80 {
81 	struct sk_buff *skb;
82 	void *reply;
83 
84 	/*
85 	 * If new attributes are added, please revisit this allocation
86 	 */
87 	skb = genlmsg_new(size, GFP_KERNEL);
88 	if (!skb)
89 		return -ENOMEM;
90 
91 	if (!info) {
92 		int seq = this_cpu_inc_return(taskstats_seqnum) - 1;
93 
94 		reply = genlmsg_put(skb, 0, seq, &family, 0, cmd);
95 	} else
96 		reply = genlmsg_put_reply(skb, info, &family, 0, cmd);
97 	if (reply == NULL) {
98 		nlmsg_free(skb);
99 		return -EINVAL;
100 	}
101 
102 	*skbp = skb;
103 	return 0;
104 }
105 
106 /*
107  * Send taskstats data in @skb to listener with nl_pid @pid
108  */
109 static int send_reply(struct sk_buff *skb, struct genl_info *info)
110 {
111 	struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
112 	void *reply = genlmsg_data(genlhdr);
113 	int rc;
114 
115 	rc = genlmsg_end(skb, reply);
116 	if (rc < 0) {
117 		nlmsg_free(skb);
118 		return rc;
119 	}
120 
121 	return genlmsg_reply(skb, info);
122 }
123 
124 /*
125  * Send taskstats data in @skb to listeners registered for @cpu's exit data
126  */
127 static void send_cpu_listeners(struct sk_buff *skb,
128 					struct listener_list *listeners)
129 {
130 	struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
131 	struct listener *s, *tmp;
132 	struct sk_buff *skb_next, *skb_cur = skb;
133 	void *reply = genlmsg_data(genlhdr);
134 	int rc, delcount = 0;
135 
136 	rc = genlmsg_end(skb, reply);
137 	if (rc < 0) {
138 		nlmsg_free(skb);
139 		return;
140 	}
141 
142 	rc = 0;
143 	down_read(&listeners->sem);
144 	list_for_each_entry(s, &listeners->list, list) {
145 		skb_next = NULL;
146 		if (!list_is_last(&s->list, &listeners->list)) {
147 			skb_next = skb_clone(skb_cur, GFP_KERNEL);
148 			if (!skb_next)
149 				break;
150 		}
151 		rc = genlmsg_unicast(&init_net, skb_cur, s->pid);
152 		if (rc == -ECONNREFUSED) {
153 			s->valid = 0;
154 			delcount++;
155 		}
156 		skb_cur = skb_next;
157 	}
158 	up_read(&listeners->sem);
159 
160 	if (skb_cur)
161 		nlmsg_free(skb_cur);
162 
163 	if (!delcount)
164 		return;
165 
166 	/* Delete invalidated entries */
167 	down_write(&listeners->sem);
168 	list_for_each_entry_safe(s, tmp, &listeners->list, list) {
169 		if (!s->valid) {
170 			list_del(&s->list);
171 			kfree(s);
172 		}
173 	}
174 	up_write(&listeners->sem);
175 }
176 
177 static void fill_stats(struct task_struct *tsk, struct taskstats *stats)
178 {
179 	memset(stats, 0, sizeof(*stats));
180 	/*
181 	 * Each accounting subsystem adds calls to its functions to
182 	 * fill in relevant parts of struct taskstsats as follows
183 	 *
184 	 *	per-task-foo(stats, tsk);
185 	 */
186 
187 	delayacct_add_tsk(stats, tsk);
188 
189 	/* fill in basic acct fields */
190 	stats->version = TASKSTATS_VERSION;
191 	stats->nvcsw = tsk->nvcsw;
192 	stats->nivcsw = tsk->nivcsw;
193 	bacct_add_tsk(stats, tsk);
194 
195 	/* fill in extended acct fields */
196 	xacct_add_tsk(stats, tsk);
197 }
198 
199 static int fill_stats_for_pid(pid_t pid, struct taskstats *stats)
200 {
201 	struct task_struct *tsk;
202 
203 	rcu_read_lock();
204 	tsk = find_task_by_vpid(pid);
205 	if (tsk)
206 		get_task_struct(tsk);
207 	rcu_read_unlock();
208 	if (!tsk)
209 		return -ESRCH;
210 	fill_stats(tsk, stats);
211 	put_task_struct(tsk);
212 	return 0;
213 }
214 
215 static int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
216 {
217 	struct task_struct *tsk, *first;
218 	unsigned long flags;
219 	int rc = -ESRCH;
220 
221 	/*
222 	 * Add additional stats from live tasks except zombie thread group
223 	 * leaders who are already counted with the dead tasks
224 	 */
225 	rcu_read_lock();
226 	first = find_task_by_vpid(tgid);
227 
228 	if (!first || !lock_task_sighand(first, &flags))
229 		goto out;
230 
231 	if (first->signal->stats)
232 		memcpy(stats, first->signal->stats, sizeof(*stats));
233 	else
234 		memset(stats, 0, sizeof(*stats));
235 
236 	tsk = first;
237 	do {
238 		if (tsk->exit_state)
239 			continue;
240 		/*
241 		 * Accounting subsystem can call its functions here to
242 		 * fill in relevant parts of struct taskstsats as follows
243 		 *
244 		 *	per-task-foo(stats, tsk);
245 		 */
246 		delayacct_add_tsk(stats, tsk);
247 
248 		stats->nvcsw += tsk->nvcsw;
249 		stats->nivcsw += tsk->nivcsw;
250 	} while_each_thread(first, tsk);
251 
252 	unlock_task_sighand(first, &flags);
253 	rc = 0;
254 out:
255 	rcu_read_unlock();
256 
257 	stats->version = TASKSTATS_VERSION;
258 	/*
259 	 * Accounting subsystems can also add calls here to modify
260 	 * fields of taskstats.
261 	 */
262 	return rc;
263 }
264 
265 static void fill_tgid_exit(struct task_struct *tsk)
266 {
267 	unsigned long flags;
268 
269 	spin_lock_irqsave(&tsk->sighand->siglock, flags);
270 	if (!tsk->signal->stats)
271 		goto ret;
272 
273 	/*
274 	 * Each accounting subsystem calls its functions here to
275 	 * accumalate its per-task stats for tsk, into the per-tgid structure
276 	 *
277 	 *	per-task-foo(tsk->signal->stats, tsk);
278 	 */
279 	delayacct_add_tsk(tsk->signal->stats, tsk);
280 ret:
281 	spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
282 	return;
283 }
284 
285 static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
286 {
287 	struct listener_list *listeners;
288 	struct listener *s, *tmp;
289 	unsigned int cpu;
290 
291 	if (!cpumask_subset(mask, cpu_possible_mask))
292 		return -EINVAL;
293 
294 	if (isadd == REGISTER) {
295 		for_each_cpu(cpu, mask) {
296 			s = kmalloc_node(sizeof(struct listener), GFP_KERNEL,
297 					 cpu_to_node(cpu));
298 			if (!s)
299 				goto cleanup;
300 			s->pid = pid;
301 			INIT_LIST_HEAD(&s->list);
302 			s->valid = 1;
303 
304 			listeners = &per_cpu(listener_array, cpu);
305 			down_write(&listeners->sem);
306 			list_add(&s->list, &listeners->list);
307 			up_write(&listeners->sem);
308 		}
309 		return 0;
310 	}
311 
312 	/* Deregister or cleanup */
313 cleanup:
314 	for_each_cpu(cpu, mask) {
315 		listeners = &per_cpu(listener_array, cpu);
316 		down_write(&listeners->sem);
317 		list_for_each_entry_safe(s, tmp, &listeners->list, list) {
318 			if (s->pid == pid) {
319 				list_del(&s->list);
320 				kfree(s);
321 				break;
322 			}
323 		}
324 		up_write(&listeners->sem);
325 	}
326 	return 0;
327 }
328 
329 static int parse(struct nlattr *na, struct cpumask *mask)
330 {
331 	char *data;
332 	int len;
333 	int ret;
334 
335 	if (na == NULL)
336 		return 1;
337 	len = nla_len(na);
338 	if (len > TASKSTATS_CPUMASK_MAXLEN)
339 		return -E2BIG;
340 	if (len < 1)
341 		return -EINVAL;
342 	data = kmalloc(len, GFP_KERNEL);
343 	if (!data)
344 		return -ENOMEM;
345 	nla_strlcpy(data, na, len);
346 	ret = cpulist_parse(data, mask);
347 	kfree(data);
348 	return ret;
349 }
350 
351 #if defined(CONFIG_64BIT) && !defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
352 #define TASKSTATS_NEEDS_PADDING 1
353 #endif
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 	/*
365 	 * The taskstats structure is internally aligned on 8 byte
366 	 * boundaries but the layout of the aggregrate reply, with
367 	 * two NLA headers and the pid (each 4 bytes), actually
368 	 * force the entire structure to be unaligned. This causes
369 	 * the kernel to issue unaligned access warnings on some
370 	 * architectures like ia64. Unfortunately, some software out there
371 	 * doesn't properly unroll the NLA packet and assumes that the start
372 	 * of the taskstats structure will always be 20 bytes from the start
373 	 * of the netlink payload. Aligning the start of the taskstats
374 	 * structure breaks this software, which we don't want. So, for now
375 	 * the alignment only happens on architectures that require it
376 	 * and those users will have to update to fixed versions of those
377 	 * packages. Space is reserved in the packet only when needed.
378 	 * This ifdef should be removed in several years e.g. 2012 once
379 	 * we can be confident that fixed versions are installed on most
380 	 * systems. We add the padding before the aggregate since the
381 	 * aggregate is already a defined type.
382 	 */
383 #ifdef TASKSTATS_NEEDS_PADDING
384 	if (nla_put(skb, TASKSTATS_TYPE_NULL, 0, NULL) < 0)
385 		goto err;
386 #endif
387 	na = nla_nest_start(skb, aggr);
388 	if (!na)
389 		goto err;
390 
391 	if (nla_put(skb, type, sizeof(pid), &pid) < 0)
392 		goto err;
393 	ret = nla_reserve(skb, TASKSTATS_TYPE_STATS, sizeof(struct taskstats));
394 	if (!ret)
395 		goto err;
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 file *file;
412 	int fput_needed;
413 
414 	na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
415 	if (!na)
416 		return -EINVAL;
417 
418 	fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
419 	file = fget_light(fd, &fput_needed);
420 	if (!file)
421 		return 0;
422 
423 	size = nla_total_size(sizeof(struct cgroupstats));
424 
425 	rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
426 				size);
427 	if (rc < 0)
428 		goto err;
429 
430 	na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
431 				sizeof(struct cgroupstats));
432 	stats = nla_data(na);
433 	memset(stats, 0, sizeof(*stats));
434 
435 	rc = cgroupstats_build(stats, file->f_dentry);
436 	if (rc < 0) {
437 		nlmsg_free(rep_skb);
438 		goto err;
439 	}
440 
441 	rc = send_reply(rep_skb, info);
442 
443 err:
444 	fput_light(file, fput_needed);
445 	return rc;
446 }
447 
448 static int cmd_attr_register_cpumask(struct genl_info *info)
449 {
450 	cpumask_var_t mask;
451 	int rc;
452 
453 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
454 		return -ENOMEM;
455 	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
456 	if (rc < 0)
457 		goto out;
458 	rc = add_del_listener(info->snd_pid, mask, REGISTER);
459 out:
460 	free_cpumask_var(mask);
461 	return rc;
462 }
463 
464 static int cmd_attr_deregister_cpumask(struct genl_info *info)
465 {
466 	cpumask_var_t mask;
467 	int rc;
468 
469 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
470 		return -ENOMEM;
471 	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
472 	if (rc < 0)
473 		goto out;
474 	rc = add_del_listener(info->snd_pid, mask, DEREGISTER);
475 out:
476 	free_cpumask_var(mask);
477 	return rc;
478 }
479 
480 static size_t taskstats_packet_size(void)
481 {
482 	size_t size;
483 
484 	size = nla_total_size(sizeof(u32)) +
485 		nla_total_size(sizeof(struct taskstats)) + nla_total_size(0);
486 #ifdef TASKSTATS_NEEDS_PADDING
487 	size += nla_total_size(0); /* Padding for alignment */
488 #endif
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 = __this_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, tsk->pid);
623 	if (!stats)
624 		goto err;
625 
626 	fill_stats(tsk, stats);
627 
628 	/*
629 	 * Doesn't matter if tsk is the leader or the last group member leaving
630 	 */
631 	if (!is_thread_group || !group_dead)
632 		goto send;
633 
634 	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tsk->tgid);
635 	if (!stats)
636 		goto err;
637 
638 	memcpy(stats, tsk->signal->stats, sizeof(*stats));
639 
640 send:
641 	send_cpu_listeners(rep_skb, listeners);
642 	return;
643 err:
644 	nlmsg_free(rep_skb);
645 }
646 
647 static struct genl_ops taskstats_ops = {
648 	.cmd		= TASKSTATS_CMD_GET,
649 	.doit		= taskstats_user_cmd,
650 	.policy		= taskstats_cmd_get_policy,
651 };
652 
653 static struct genl_ops cgroupstats_ops = {
654 	.cmd		= CGROUPSTATS_CMD_GET,
655 	.doit		= cgroupstats_user_cmd,
656 	.policy		= cgroupstats_cmd_get_policy,
657 };
658 
659 /* Needed early in initialization */
660 void __init taskstats_init_early(void)
661 {
662 	unsigned int i;
663 
664 	taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
665 	for_each_possible_cpu(i) {
666 		INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
667 		init_rwsem(&(per_cpu(listener_array, i).sem));
668 	}
669 }
670 
671 static int __init taskstats_init(void)
672 {
673 	int rc;
674 
675 	rc = genl_register_family(&family);
676 	if (rc)
677 		return rc;
678 
679 	rc = genl_register_ops(&family, &taskstats_ops);
680 	if (rc < 0)
681 		goto err;
682 
683 	rc = genl_register_ops(&family, &cgroupstats_ops);
684 	if (rc < 0)
685 		goto err_cgroup_ops;
686 
687 	family_registered = 1;
688 	printk("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
689 	return 0;
690 err_cgroup_ops:
691 	genl_unregister_ops(&family, &taskstats_ops);
692 err:
693 	genl_unregister_family(&family);
694 	return rc;
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