taskstats.c revision 1a51410abe7d0ee4b1d112780f46df87d3621043 
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 <linux/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
39static DEFINE_PER_CPU(__u32, taskstats_seqnum);
40static int family_registered;
41struct kmem_cache *taskstats_cache;
42
43static 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
50static 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
56static const struct nla_policy cgroupstats_cmd_get_policy[CGROUPSTATS_CMD_ATTR_MAX+1] = {
57	[CGROUPSTATS_CMD_ATTR_FD] = { .type = NLA_U32 },
58};
59
60struct listener {
61	struct list_head list;
62	pid_t pid;
63	char valid;
64};
65
66struct listener_list {
67	struct rw_semaphore sem;
68	struct list_head list;
69};
70static DEFINE_PER_CPU(struct listener_list, listener_array);
71
72enum actions {
73	REGISTER,
74	DEREGISTER,
75	CPU_DONT_CARE
76};
77
78static 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 */
109static 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 */
127static 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
177static 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
199static 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
215static 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;
254out:
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
265static 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);
280ret:
281	spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
282	return;
283}
284
285static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
286{
287	struct listener_list *listeners;
288	struct listener *s, *tmp, *s2;
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),
297					GFP_KERNEL, cpu_to_node(cpu));
298			if (!s)
299				goto cleanup;
300
301			s->pid = pid;
302			s->valid = 1;
303
304			listeners = &per_cpu(listener_array, cpu);
305			down_write(&listeners->sem);
306			list_for_each_entry(s2, &listeners->list, list) {
307				if (s2->pid == pid && s2->valid)
308					goto exists;
309			}
310			list_add(&s->list, &listeners->list);
311			s = NULL;
312exists:
313			up_write(&listeners->sem);
314			kfree(s); /* nop if NULL */
315		}
316		return 0;
317	}
318
319	/* Deregister or cleanup */
320cleanup:
321	for_each_cpu(cpu, mask) {
322		listeners = &per_cpu(listener_array, cpu);
323		down_write(&listeners->sem);
324		list_for_each_entry_safe(s, tmp, &listeners->list, list) {
325			if (s->pid == pid) {
326				list_del(&s->list);
327				kfree(s);
328				break;
329			}
330		}
331		up_write(&listeners->sem);
332	}
333	return 0;
334}
335
336static int parse(struct nlattr *na, struct cpumask *mask)
337{
338	char *data;
339	int len;
340	int ret;
341
342	if (na == NULL)
343		return 1;
344	len = nla_len(na);
345	if (len > TASKSTATS_CPUMASK_MAXLEN)
346		return -E2BIG;
347	if (len < 1)
348		return -EINVAL;
349	data = kmalloc(len, GFP_KERNEL);
350	if (!data)
351		return -ENOMEM;
352	nla_strlcpy(data, na, len);
353	ret = cpulist_parse(data, mask);
354	kfree(data);
355	return ret;
356}
357
358#if defined(CONFIG_64BIT) && !defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
359#define TASKSTATS_NEEDS_PADDING 1
360#endif
361
362static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
363{
364	struct nlattr *na, *ret;
365	int aggr;
366
367	aggr = (type == TASKSTATS_TYPE_PID)
368			? TASKSTATS_TYPE_AGGR_PID
369			: TASKSTATS_TYPE_AGGR_TGID;
370
371	/*
372	 * The taskstats structure is internally aligned on 8 byte
373	 * boundaries but the layout of the aggregrate reply, with
374	 * two NLA headers and the pid (each 4 bytes), actually
375	 * force the entire structure to be unaligned. This causes
376	 * the kernel to issue unaligned access warnings on some
377	 * architectures like ia64. Unfortunately, some software out there
378	 * doesn't properly unroll the NLA packet and assumes that the start
379	 * of the taskstats structure will always be 20 bytes from the start
380	 * of the netlink payload. Aligning the start of the taskstats
381	 * structure breaks this software, which we don't want. So, for now
382	 * the alignment only happens on architectures that require it
383	 * and those users will have to update to fixed versions of those
384	 * packages. Space is reserved in the packet only when needed.
385	 * This ifdef should be removed in several years e.g. 2012 once
386	 * we can be confident that fixed versions are installed on most
387	 * systems. We add the padding before the aggregate since the
388	 * aggregate is already a defined type.
389	 */
390#ifdef TASKSTATS_NEEDS_PADDING
391	if (nla_put(skb, TASKSTATS_TYPE_NULL, 0, NULL) < 0)
392		goto err;
393#endif
394	na = nla_nest_start(skb, aggr);
395	if (!na)
396		goto err;
397
398	if (nla_put(skb, type, sizeof(pid), &pid) < 0)
399		goto err;
400	ret = nla_reserve(skb, TASKSTATS_TYPE_STATS, sizeof(struct taskstats));
401	if (!ret)
402		goto err;
403	nla_nest_end(skb, na);
404
405	return nla_data(ret);
406err:
407	return NULL;
408}
409
410static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
411{
412	int rc = 0;
413	struct sk_buff *rep_skb;
414	struct cgroupstats *stats;
415	struct nlattr *na;
416	size_t size;
417	u32 fd;
418	struct file *file;
419	int fput_needed;
420
421	na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
422	if (!na)
423		return -EINVAL;
424
425	fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
426	file = fget_light(fd, &fput_needed);
427	if (!file)
428		return 0;
429
430	size = nla_total_size(sizeof(struct cgroupstats));
431
432	rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
433				size);
434	if (rc < 0)
435		goto err;
436
437	na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
438				sizeof(struct cgroupstats));
439	stats = nla_data(na);
440	memset(stats, 0, sizeof(*stats));
441
442	rc = cgroupstats_build(stats, file->f_dentry);
443	if (rc < 0) {
444		nlmsg_free(rep_skb);
445		goto err;
446	}
447
448	rc = send_reply(rep_skb, info);
449
450err:
451	fput_light(file, fput_needed);
452	return rc;
453}
454
455static int cmd_attr_register_cpumask(struct genl_info *info)
456{
457	cpumask_var_t mask;
458	int rc;
459
460	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
461		return -ENOMEM;
462	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
463	if (rc < 0)
464		goto out;
465	rc = add_del_listener(info->snd_pid, mask, REGISTER);
466out:
467	free_cpumask_var(mask);
468	return rc;
469}
470
471static int cmd_attr_deregister_cpumask(struct genl_info *info)
472{
473	cpumask_var_t mask;
474	int rc;
475
476	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
477		return -ENOMEM;
478	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
479	if (rc < 0)
480		goto out;
481	rc = add_del_listener(info->snd_pid, mask, DEREGISTER);
482out:
483	free_cpumask_var(mask);
484	return rc;
485}
486
487static size_t taskstats_packet_size(void)
488{
489	size_t size;
490
491	size = nla_total_size(sizeof(u32)) +
492		nla_total_size(sizeof(struct taskstats)) + nla_total_size(0);
493#ifdef TASKSTATS_NEEDS_PADDING
494	size += nla_total_size(0); /* Padding for alignment */
495#endif
496	return size;
497}
498
499static int cmd_attr_pid(struct genl_info *info)
500{
501	struct taskstats *stats;
502	struct sk_buff *rep_skb;
503	size_t size;
504	u32 pid;
505	int rc;
506
507	size = taskstats_packet_size();
508
509	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
510	if (rc < 0)
511		return rc;
512
513	rc = -EINVAL;
514	pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
515	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
516	if (!stats)
517		goto err;
518
519	rc = fill_stats_for_pid(pid, stats);
520	if (rc < 0)
521		goto err;
522	return send_reply(rep_skb, info);
523err:
524	nlmsg_free(rep_skb);
525	return rc;
526}
527
528static int cmd_attr_tgid(struct genl_info *info)
529{
530	struct taskstats *stats;
531	struct sk_buff *rep_skb;
532	size_t size;
533	u32 tgid;
534	int rc;
535
536	size = taskstats_packet_size();
537
538	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
539	if (rc < 0)
540		return rc;
541
542	rc = -EINVAL;
543	tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
544	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
545	if (!stats)
546		goto err;
547
548	rc = fill_stats_for_tgid(tgid, stats);
549	if (rc < 0)
550		goto err;
551	return send_reply(rep_skb, info);
552err:
553	nlmsg_free(rep_skb);
554	return rc;
555}
556
557static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
558{
559	if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
560		return cmd_attr_register_cpumask(info);
561	else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
562		return cmd_attr_deregister_cpumask(info);
563	else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
564		return cmd_attr_pid(info);
565	else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
566		return cmd_attr_tgid(info);
567	else
568		return -EINVAL;
569}
570
571static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
572{
573	struct signal_struct *sig = tsk->signal;
574	struct taskstats *stats;
575
576	if (sig->stats || thread_group_empty(tsk))
577		goto ret;
578
579	/* No problem if kmem_cache_zalloc() fails */
580	stats = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
581
582	spin_lock_irq(&tsk->sighand->siglock);
583	if (!sig->stats) {
584		sig->stats = stats;
585		stats = NULL;
586	}
587	spin_unlock_irq(&tsk->sighand->siglock);
588
589	if (stats)
590		kmem_cache_free(taskstats_cache, stats);
591ret:
592	return sig->stats;
593}
594
595/* Send pid data out on exit */
596void taskstats_exit(struct task_struct *tsk, int group_dead)
597{
598	int rc;
599	struct listener_list *listeners;
600	struct taskstats *stats;
601	struct sk_buff *rep_skb;
602	size_t size;
603	int is_thread_group;
604
605	if (!family_registered)
606		return;
607
608	/*
609	 * Size includes space for nested attributes
610	 */
611	size = taskstats_packet_size();
612
613	is_thread_group = !!taskstats_tgid_alloc(tsk);
614	if (is_thread_group) {
615		/* PID + STATS + TGID + STATS */
616		size = 2 * size;
617		/* fill the tsk->signal->stats structure */
618		fill_tgid_exit(tsk);
619	}
620
621	listeners = __this_cpu_ptr(&listener_array);
622	if (list_empty(&listeners->list))
623		return;
624
625	rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
626	if (rc < 0)
627		return;
628
629	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, tsk->pid);
630	if (!stats)
631		goto err;
632
633	fill_stats(tsk, stats);
634
635	/*
636	 * Doesn't matter if tsk is the leader or the last group member leaving
637	 */
638	if (!is_thread_group || !group_dead)
639		goto send;
640
641	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tsk->tgid);
642	if (!stats)
643		goto err;
644
645	memcpy(stats, tsk->signal->stats, sizeof(*stats));
646
647send:
648	send_cpu_listeners(rep_skb, listeners);
649	return;
650err:
651	nlmsg_free(rep_skb);
652}
653
654static struct genl_ops taskstats_ops = {
655	.cmd		= TASKSTATS_CMD_GET,
656	.doit		= taskstats_user_cmd,
657	.policy		= taskstats_cmd_get_policy,
658	.flags		= GENL_ADMIN_PERM,
659};
660
661static struct genl_ops cgroupstats_ops = {
662	.cmd		= CGROUPSTATS_CMD_GET,
663	.doit		= cgroupstats_user_cmd,
664	.policy		= cgroupstats_cmd_get_policy,
665};
666
667/* Needed early in initialization */
668void __init taskstats_init_early(void)
669{
670	unsigned int i;
671
672	taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
673	for_each_possible_cpu(i) {
674		INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
675		init_rwsem(&(per_cpu(listener_array, i).sem));
676	}
677}
678
679static int __init taskstats_init(void)
680{
681	int rc;
682
683	rc = genl_register_family(&family);
684	if (rc)
685		return rc;
686
687	rc = genl_register_ops(&family, &taskstats_ops);
688	if (rc < 0)
689		goto err;
690
691	rc = genl_register_ops(&family, &cgroupstats_ops);
692	if (rc < 0)
693		goto err_cgroup_ops;
694
695	family_registered = 1;
696	pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
697	return 0;
698err_cgroup_ops:
699	genl_unregister_ops(&family, &taskstats_ops);
700err:
701	genl_unregister_family(&family);
702	return rc;
703}
704
705/*
706 * late initcall ensures initialization of statistics collection
707 * mechanisms precedes initialization of the taskstats interface
708 */
709late_initcall(taskstats_init);
710