taskstats.c revision 4bd6e32acec66c55c6c1af4672f3216b2ac88e35 
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
34/*
35 * Maximum length of a cpumask that can be specified in
36 * the TASKSTATS_CMD_ATTR_REGISTER/DEREGISTER_CPUMASK attribute
37 */
38#define TASKSTATS_CPUMASK_MAXLEN	(100+6*NR_CPUS)
39
40static DEFINE_PER_CPU(__u32, taskstats_seqnum);
41static int family_registered;
42struct kmem_cache *taskstats_cache;
43
44static struct genl_family family = {
45	.id		= GENL_ID_GENERATE,
46	.name		= TASKSTATS_GENL_NAME,
47	.version	= TASKSTATS_GENL_VERSION,
48	.maxattr	= TASKSTATS_CMD_ATTR_MAX,
49};
50
51static const struct nla_policy taskstats_cmd_get_policy[TASKSTATS_CMD_ATTR_MAX+1] = {
52	[TASKSTATS_CMD_ATTR_PID]  = { .type = NLA_U32 },
53	[TASKSTATS_CMD_ATTR_TGID] = { .type = NLA_U32 },
54	[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK] = { .type = NLA_STRING },
55	[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK] = { .type = NLA_STRING },};
56
57static const struct nla_policy cgroupstats_cmd_get_policy[CGROUPSTATS_CMD_ATTR_MAX+1] = {
58	[CGROUPSTATS_CMD_ATTR_FD] = { .type = NLA_U32 },
59};
60
61struct listener {
62	struct list_head list;
63	pid_t pid;
64	char valid;
65};
66
67struct listener_list {
68	struct rw_semaphore sem;
69	struct list_head list;
70};
71static DEFINE_PER_CPU(struct listener_list, listener_array);
72
73enum actions {
74	REGISTER,
75	DEREGISTER,
76	CPU_DONT_CARE
77};
78
79static 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 */
110static 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	int rc;
115
116	rc = genlmsg_end(skb, reply);
117	if (rc < 0) {
118		nlmsg_free(skb);
119		return rc;
120	}
121
122	return genlmsg_reply(skb, info);
123}
124
125/*
126 * Send taskstats data in @skb to listeners registered for @cpu's exit data
127 */
128static void send_cpu_listeners(struct sk_buff *skb,
129					struct listener_list *listeners)
130{
131	struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
132	struct listener *s, *tmp;
133	struct sk_buff *skb_next, *skb_cur = skb;
134	void *reply = genlmsg_data(genlhdr);
135	int rc, delcount = 0;
136
137	rc = genlmsg_end(skb, reply);
138	if (rc < 0) {
139		nlmsg_free(skb);
140		return;
141	}
142
143	rc = 0;
144	down_read(&listeners->sem);
145	list_for_each_entry(s, &listeners->list, list) {
146		skb_next = NULL;
147		if (!list_is_last(&s->list, &listeners->list)) {
148			skb_next = skb_clone(skb_cur, GFP_KERNEL);
149			if (!skb_next)
150				break;
151		}
152		rc = genlmsg_unicast(&init_net, skb_cur, s->pid);
153		if (rc == -ECONNREFUSED) {
154			s->valid = 0;
155			delcount++;
156		}
157		skb_cur = skb_next;
158	}
159	up_read(&listeners->sem);
160
161	if (skb_cur)
162		nlmsg_free(skb_cur);
163
164	if (!delcount)
165		return;
166
167	/* Delete invalidated entries */
168	down_write(&listeners->sem);
169	list_for_each_entry_safe(s, tmp, &listeners->list, list) {
170		if (!s->valid) {
171			list_del(&s->list);
172			kfree(s);
173		}
174	}
175	up_write(&listeners->sem);
176}
177
178static void fill_stats(struct user_namespace *user_ns,
179		       struct pid_namespace *pid_ns,
180		       struct task_struct *tsk, struct taskstats *stats)
181{
182	memset(stats, 0, sizeof(*stats));
183	/*
184	 * Each accounting subsystem adds calls to its functions to
185	 * fill in relevant parts of struct taskstsats as follows
186	 *
187	 *	per-task-foo(stats, tsk);
188	 */
189
190	delayacct_add_tsk(stats, tsk);
191
192	/* fill in basic acct fields */
193	stats->version = TASKSTATS_VERSION;
194	stats->nvcsw = tsk->nvcsw;
195	stats->nivcsw = tsk->nivcsw;
196	bacct_add_tsk(user_ns, pid_ns, stats, tsk);
197
198	/* fill in extended acct fields */
199	xacct_add_tsk(stats, tsk);
200}
201
202static int fill_stats_for_pid(pid_t pid, struct taskstats *stats)
203{
204	struct task_struct *tsk;
205
206	rcu_read_lock();
207	tsk = find_task_by_vpid(pid);
208	if (tsk)
209		get_task_struct(tsk);
210	rcu_read_unlock();
211	if (!tsk)
212		return -ESRCH;
213	fill_stats(current_user_ns(), task_active_pid_ns(current), tsk, stats);
214	put_task_struct(tsk);
215	return 0;
216}
217
218static int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
219{
220	struct task_struct *tsk, *first;
221	unsigned long flags;
222	int rc = -ESRCH;
223
224	/*
225	 * Add additional stats from live tasks except zombie thread group
226	 * leaders who are already counted with the dead tasks
227	 */
228	rcu_read_lock();
229	first = find_task_by_vpid(tgid);
230
231	if (!first || !lock_task_sighand(first, &flags))
232		goto out;
233
234	if (first->signal->stats)
235		memcpy(stats, first->signal->stats, sizeof(*stats));
236	else
237		memset(stats, 0, sizeof(*stats));
238
239	tsk = first;
240	do {
241		if (tsk->exit_state)
242			continue;
243		/*
244		 * Accounting subsystem can call its functions here to
245		 * fill in relevant parts of struct taskstsats as follows
246		 *
247		 *	per-task-foo(stats, tsk);
248		 */
249		delayacct_add_tsk(stats, tsk);
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;
257out:
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
268static 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);
283ret:
284	spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
285	return;
286}
287
288static 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
294	if (!cpumask_subset(mask, cpu_possible_mask))
295		return -EINVAL;
296
297	if (current_user_ns() != &init_user_ns)
298		return -EINVAL;
299
300	if (task_active_pid_ns(current) != &init_pid_ns)
301		return -EINVAL;
302
303	if (isadd == REGISTER) {
304		for_each_cpu(cpu, mask) {
305			s = kmalloc_node(sizeof(struct listener),
306					GFP_KERNEL, cpu_to_node(cpu));
307			if (!s)
308				goto cleanup;
309
310			s->pid = pid;
311			s->valid = 1;
312
313			listeners = &per_cpu(listener_array, cpu);
314			down_write(&listeners->sem);
315			list_for_each_entry(s2, &listeners->list, list) {
316				if (s2->pid == pid && s2->valid)
317					goto exists;
318			}
319			list_add(&s->list, &listeners->list);
320			s = NULL;
321exists:
322			up_write(&listeners->sem);
323			kfree(s); /* nop if NULL */
324		}
325		return 0;
326	}
327
328	/* Deregister or cleanup */
329cleanup:
330	for_each_cpu(cpu, mask) {
331		listeners = &per_cpu(listener_array, cpu);
332		down_write(&listeners->sem);
333		list_for_each_entry_safe(s, tmp, &listeners->list, list) {
334			if (s->pid == pid) {
335				list_del(&s->list);
336				kfree(s);
337				break;
338			}
339		}
340		up_write(&listeners->sem);
341	}
342	return 0;
343}
344
345static int parse(struct nlattr *na, struct cpumask *mask)
346{
347	char *data;
348	int len;
349	int ret;
350
351	if (na == NULL)
352		return 1;
353	len = nla_len(na);
354	if (len > TASKSTATS_CPUMASK_MAXLEN)
355		return -E2BIG;
356	if (len < 1)
357		return -EINVAL;
358	data = kmalloc(len, GFP_KERNEL);
359	if (!data)
360		return -ENOMEM;
361	nla_strlcpy(data, na, len);
362	ret = cpulist_parse(data, mask);
363	kfree(data);
364	return ret;
365}
366
367#if defined(CONFIG_64BIT) && !defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
368#define TASKSTATS_NEEDS_PADDING 1
369#endif
370
371static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
372{
373	struct nlattr *na, *ret;
374	int aggr;
375
376	aggr = (type == TASKSTATS_TYPE_PID)
377			? TASKSTATS_TYPE_AGGR_PID
378			: TASKSTATS_TYPE_AGGR_TGID;
379
380	/*
381	 * The taskstats structure is internally aligned on 8 byte
382	 * boundaries but the layout of the aggregrate reply, with
383	 * two NLA headers and the pid (each 4 bytes), actually
384	 * force the entire structure to be unaligned. This causes
385	 * the kernel to issue unaligned access warnings on some
386	 * architectures like ia64. Unfortunately, some software out there
387	 * doesn't properly unroll the NLA packet and assumes that the start
388	 * of the taskstats structure will always be 20 bytes from the start
389	 * of the netlink payload. Aligning the start of the taskstats
390	 * structure breaks this software, which we don't want. So, for now
391	 * the alignment only happens on architectures that require it
392	 * and those users will have to update to fixed versions of those
393	 * packages. Space is reserved in the packet only when needed.
394	 * This ifdef should be removed in several years e.g. 2012 once
395	 * we can be confident that fixed versions are installed on most
396	 * systems. We add the padding before the aggregate since the
397	 * aggregate is already a defined type.
398	 */
399#ifdef TASKSTATS_NEEDS_PADDING
400	if (nla_put(skb, TASKSTATS_TYPE_NULL, 0, NULL) < 0)
401		goto err;
402#endif
403	na = nla_nest_start(skb, aggr);
404	if (!na)
405		goto err;
406
407	if (nla_put(skb, type, sizeof(pid), &pid) < 0)
408		goto err;
409	ret = nla_reserve(skb, TASKSTATS_TYPE_STATS, sizeof(struct taskstats));
410	if (!ret)
411		goto err;
412	nla_nest_end(skb, na);
413
414	return nla_data(ret);
415err:
416	return NULL;
417}
418
419static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
420{
421	int rc = 0;
422	struct sk_buff *rep_skb;
423	struct cgroupstats *stats;
424	struct nlattr *na;
425	size_t size;
426	u32 fd;
427	struct file *file;
428	int fput_needed;
429
430	na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
431	if (!na)
432		return -EINVAL;
433
434	fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
435	file = fget_light(fd, &fput_needed);
436	if (!file)
437		return 0;
438
439	size = nla_total_size(sizeof(struct cgroupstats));
440
441	rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
442				size);
443	if (rc < 0)
444		goto err;
445
446	na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
447				sizeof(struct cgroupstats));
448	if (na == NULL) {
449		rc = -EMSGSIZE;
450		goto err;
451	}
452
453	stats = nla_data(na);
454	memset(stats, 0, sizeof(*stats));
455
456	rc = cgroupstats_build(stats, file->f_dentry);
457	if (rc < 0) {
458		nlmsg_free(rep_skb);
459		goto err;
460	}
461
462	rc = send_reply(rep_skb, info);
463
464err:
465	fput_light(file, fput_needed);
466	return rc;
467}
468
469static int cmd_attr_register_cpumask(struct genl_info *info)
470{
471	cpumask_var_t mask;
472	int rc;
473
474	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
475		return -ENOMEM;
476	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
477	if (rc < 0)
478		goto out;
479	rc = add_del_listener(info->snd_pid, mask, REGISTER);
480out:
481	free_cpumask_var(mask);
482	return rc;
483}
484
485static int cmd_attr_deregister_cpumask(struct genl_info *info)
486{
487	cpumask_var_t mask;
488	int rc;
489
490	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
491		return -ENOMEM;
492	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
493	if (rc < 0)
494		goto out;
495	rc = add_del_listener(info->snd_pid, mask, DEREGISTER);
496out:
497	free_cpumask_var(mask);
498	return rc;
499}
500
501static size_t taskstats_packet_size(void)
502{
503	size_t size;
504
505	size = nla_total_size(sizeof(u32)) +
506		nla_total_size(sizeof(struct taskstats)) + nla_total_size(0);
507#ifdef TASKSTATS_NEEDS_PADDING
508	size += nla_total_size(0); /* Padding for alignment */
509#endif
510	return size;
511}
512
513static int cmd_attr_pid(struct genl_info *info)
514{
515	struct taskstats *stats;
516	struct sk_buff *rep_skb;
517	size_t size;
518	u32 pid;
519	int rc;
520
521	size = taskstats_packet_size();
522
523	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
524	if (rc < 0)
525		return rc;
526
527	rc = -EINVAL;
528	pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
529	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
530	if (!stats)
531		goto err;
532
533	rc = fill_stats_for_pid(pid, stats);
534	if (rc < 0)
535		goto err;
536	return send_reply(rep_skb, info);
537err:
538	nlmsg_free(rep_skb);
539	return rc;
540}
541
542static int cmd_attr_tgid(struct genl_info *info)
543{
544	struct taskstats *stats;
545	struct sk_buff *rep_skb;
546	size_t size;
547	u32 tgid;
548	int rc;
549
550	size = taskstats_packet_size();
551
552	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
553	if (rc < 0)
554		return rc;
555
556	rc = -EINVAL;
557	tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
558	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
559	if (!stats)
560		goto err;
561
562	rc = fill_stats_for_tgid(tgid, stats);
563	if (rc < 0)
564		goto err;
565	return send_reply(rep_skb, info);
566err:
567	nlmsg_free(rep_skb);
568	return rc;
569}
570
571static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
572{
573	if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
574		return cmd_attr_register_cpumask(info);
575	else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
576		return cmd_attr_deregister_cpumask(info);
577	else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
578		return cmd_attr_pid(info);
579	else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
580		return cmd_attr_tgid(info);
581	else
582		return -EINVAL;
583}
584
585static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
586{
587	struct signal_struct *sig = tsk->signal;
588	struct taskstats *stats;
589
590	if (sig->stats || thread_group_empty(tsk))
591		goto ret;
592
593	/* No problem if kmem_cache_zalloc() fails */
594	stats = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
595
596	spin_lock_irq(&tsk->sighand->siglock);
597	if (!sig->stats) {
598		sig->stats = stats;
599		stats = NULL;
600	}
601	spin_unlock_irq(&tsk->sighand->siglock);
602
603	if (stats)
604		kmem_cache_free(taskstats_cache, stats);
605ret:
606	return sig->stats;
607}
608
609/* Send pid data out on exit */
610void taskstats_exit(struct task_struct *tsk, int group_dead)
611{
612	int rc;
613	struct listener_list *listeners;
614	struct taskstats *stats;
615	struct sk_buff *rep_skb;
616	size_t size;
617	int is_thread_group;
618
619	if (!family_registered)
620		return;
621
622	/*
623	 * Size includes space for nested attributes
624	 */
625	size = taskstats_packet_size();
626
627	is_thread_group = !!taskstats_tgid_alloc(tsk);
628	if (is_thread_group) {
629		/* PID + STATS + TGID + STATS */
630		size = 2 * size;
631		/* fill the tsk->signal->stats structure */
632		fill_tgid_exit(tsk);
633	}
634
635	listeners = __this_cpu_ptr(&listener_array);
636	if (list_empty(&listeners->list))
637		return;
638
639	rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
640	if (rc < 0)
641		return;
642
643	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
644			 task_pid_nr_ns(tsk, &init_pid_ns));
645	if (!stats)
646		goto err;
647
648	fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
649
650	/*
651	 * Doesn't matter if tsk is the leader or the last group member leaving
652	 */
653	if (!is_thread_group || !group_dead)
654		goto send;
655
656	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
657			 task_tgid_nr_ns(tsk, &init_pid_ns));
658	if (!stats)
659		goto err;
660
661	memcpy(stats, tsk->signal->stats, sizeof(*stats));
662
663send:
664	send_cpu_listeners(rep_skb, listeners);
665	return;
666err:
667	nlmsg_free(rep_skb);
668}
669
670static struct genl_ops taskstats_ops = {
671	.cmd		= TASKSTATS_CMD_GET,
672	.doit		= taskstats_user_cmd,
673	.policy		= taskstats_cmd_get_policy,
674	.flags		= GENL_ADMIN_PERM,
675};
676
677static struct genl_ops cgroupstats_ops = {
678	.cmd		= CGROUPSTATS_CMD_GET,
679	.doit		= cgroupstats_user_cmd,
680	.policy		= cgroupstats_cmd_get_policy,
681};
682
683/* Needed early in initialization */
684void __init taskstats_init_early(void)
685{
686	unsigned int i;
687
688	taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
689	for_each_possible_cpu(i) {
690		INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
691		init_rwsem(&(per_cpu(listener_array, i).sem));
692	}
693}
694
695static int __init taskstats_init(void)
696{
697	int rc;
698
699	rc = genl_register_family(&family);
700	if (rc)
701		return rc;
702
703	rc = genl_register_ops(&family, &taskstats_ops);
704	if (rc < 0)
705		goto err;
706
707	rc = genl_register_ops(&family, &cgroupstats_ops);
708	if (rc < 0)
709		goto err_cgroup_ops;
710
711	family_registered = 1;
712	pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
713	return 0;
714err_cgroup_ops:
715	genl_unregister_ops(&family, &taskstats_ops);
716err:
717	genl_unregister_family(&family);
718	return rc;
719}
720
721/*
722 * late initcall ensures initialization of statistics collection
723 * mechanisms precedes initialization of the taskstats interface
724 */
725late_initcall(taskstats_init);
726