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
115	genlmsg_end(skb, reply);
116
117	return genlmsg_reply(skb, info);
118}
119
120/*
121 * Send taskstats data in @skb to listeners registered for @cpu's exit data
122 */
123static void send_cpu_listeners(struct sk_buff *skb,
124					struct listener_list *listeners)
125{
126	struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
127	struct listener *s, *tmp;
128	struct sk_buff *skb_next, *skb_cur = skb;
129	void *reply = genlmsg_data(genlhdr);
130	int rc, delcount = 0;
131
132	genlmsg_end(skb, reply);
133
134	rc = 0;
135	down_read(&listeners->sem);
136	list_for_each_entry(s, &listeners->list, list) {
137		skb_next = NULL;
138		if (!list_is_last(&s->list, &listeners->list)) {
139			skb_next = skb_clone(skb_cur, GFP_KERNEL);
140			if (!skb_next)
141				break;
142		}
143		rc = genlmsg_unicast(&init_net, skb_cur, s->pid);
144		if (rc == -ECONNREFUSED) {
145			s->valid = 0;
146			delcount++;
147		}
148		skb_cur = skb_next;
149	}
150	up_read(&listeners->sem);
151
152	if (skb_cur)
153		nlmsg_free(skb_cur);
154
155	if (!delcount)
156		return;
157
158	/* Delete invalidated entries */
159	down_write(&listeners->sem);
160	list_for_each_entry_safe(s, tmp, &listeners->list, list) {
161		if (!s->valid) {
162			list_del(&s->list);
163			kfree(s);
164		}
165	}
166	up_write(&listeners->sem);
167}
168
169static void fill_stats(struct user_namespace *user_ns,
170		       struct pid_namespace *pid_ns,
171		       struct task_struct *tsk, struct taskstats *stats)
172{
173	memset(stats, 0, sizeof(*stats));
174	/*
175	 * Each accounting subsystem adds calls to its functions to
176	 * fill in relevant parts of struct taskstsats as follows
177	 *
178	 *	per-task-foo(stats, tsk);
179	 */
180
181	delayacct_add_tsk(stats, tsk);
182
183	/* fill in basic acct fields */
184	stats->version = TASKSTATS_VERSION;
185	stats->nvcsw = tsk->nvcsw;
186	stats->nivcsw = tsk->nivcsw;
187	bacct_add_tsk(user_ns, pid_ns, stats, tsk);
188
189	/* fill in extended acct fields */
190	xacct_add_tsk(stats, tsk);
191}
192
193static int fill_stats_for_pid(pid_t pid, struct taskstats *stats)
194{
195	struct task_struct *tsk;
196
197	rcu_read_lock();
198	tsk = find_task_by_vpid(pid);
199	if (tsk)
200		get_task_struct(tsk);
201	rcu_read_unlock();
202	if (!tsk)
203		return -ESRCH;
204	fill_stats(current_user_ns(), task_active_pid_ns(current), tsk, stats);
205	put_task_struct(tsk);
206	return 0;
207}
208
209static int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
210{
211	struct task_struct *tsk, *first;
212	unsigned long flags;
213	int rc = -ESRCH;
214
215	/*
216	 * Add additional stats from live tasks except zombie thread group
217	 * leaders who are already counted with the dead tasks
218	 */
219	rcu_read_lock();
220	first = find_task_by_vpid(tgid);
221
222	if (!first || !lock_task_sighand(first, &flags))
223		goto out;
224
225	if (first->signal->stats)
226		memcpy(stats, first->signal->stats, sizeof(*stats));
227	else
228		memset(stats, 0, sizeof(*stats));
229
230	tsk = first;
231	do {
232		if (tsk->exit_state)
233			continue;
234		/*
235		 * Accounting subsystem can call its functions here to
236		 * fill in relevant parts of struct taskstsats as follows
237		 *
238		 *	per-task-foo(stats, tsk);
239		 */
240		delayacct_add_tsk(stats, tsk);
241
242		stats->nvcsw += tsk->nvcsw;
243		stats->nivcsw += tsk->nivcsw;
244	} while_each_thread(first, tsk);
245
246	unlock_task_sighand(first, &flags);
247	rc = 0;
248out:
249	rcu_read_unlock();
250
251	stats->version = TASKSTATS_VERSION;
252	/*
253	 * Accounting subsystems can also add calls here to modify
254	 * fields of taskstats.
255	 */
256	return rc;
257}
258
259static void fill_tgid_exit(struct task_struct *tsk)
260{
261	unsigned long flags;
262
263	spin_lock_irqsave(&tsk->sighand->siglock, flags);
264	if (!tsk->signal->stats)
265		goto ret;
266
267	/*
268	 * Each accounting subsystem calls its functions here to
269	 * accumalate its per-task stats for tsk, into the per-tgid structure
270	 *
271	 *	per-task-foo(tsk->signal->stats, tsk);
272	 */
273	delayacct_add_tsk(tsk->signal->stats, tsk);
274ret:
275	spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
276	return;
277}
278
279static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
280{
281	struct listener_list *listeners;
282	struct listener *s, *tmp, *s2;
283	unsigned int cpu;
284	int ret = 0;
285
286	if (!cpumask_subset(mask, cpu_possible_mask))
287		return -EINVAL;
288
289	if (current_user_ns() != &init_user_ns)
290		return -EINVAL;
291
292	if (task_active_pid_ns(current) != &init_pid_ns)
293		return -EINVAL;
294
295	if (isadd == REGISTER) {
296		for_each_cpu(cpu, mask) {
297			s = kmalloc_node(sizeof(struct listener),
298					GFP_KERNEL, cpu_to_node(cpu));
299			if (!s) {
300				ret = -ENOMEM;
301				goto cleanup;
302			}
303			s->pid = pid;
304			s->valid = 1;
305
306			listeners = &per_cpu(listener_array, cpu);
307			down_write(&listeners->sem);
308			list_for_each_entry(s2, &listeners->list, list) {
309				if (s2->pid == pid && s2->valid)
310					goto exists;
311			}
312			list_add(&s->list, &listeners->list);
313			s = NULL;
314exists:
315			up_write(&listeners->sem);
316			kfree(s); /* nop if NULL */
317		}
318		return 0;
319	}
320
321	/* Deregister or cleanup */
322cleanup:
323	for_each_cpu(cpu, mask) {
324		listeners = &per_cpu(listener_array, cpu);
325		down_write(&listeners->sem);
326		list_for_each_entry_safe(s, tmp, &listeners->list, list) {
327			if (s->pid == pid) {
328				list_del(&s->list);
329				kfree(s);
330				break;
331			}
332		}
333		up_write(&listeners->sem);
334	}
335	return ret;
336}
337
338static int parse(struct nlattr *na, struct cpumask *mask)
339{
340	char *data;
341	int len;
342	int ret;
343
344	if (na == NULL)
345		return 1;
346	len = nla_len(na);
347	if (len > TASKSTATS_CPUMASK_MAXLEN)
348		return -E2BIG;
349	if (len < 1)
350		return -EINVAL;
351	data = kmalloc(len, GFP_KERNEL);
352	if (!data)
353		return -ENOMEM;
354	nla_strlcpy(data, na, len);
355	ret = cpulist_parse(data, mask);
356	kfree(data);
357	return ret;
358}
359
360#if defined(CONFIG_64BIT) && !defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
361#define TASKSTATS_NEEDS_PADDING 1
362#endif
363
364static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
365{
366	struct nlattr *na, *ret;
367	int aggr;
368
369	aggr = (type == TASKSTATS_TYPE_PID)
370			? TASKSTATS_TYPE_AGGR_PID
371			: TASKSTATS_TYPE_AGGR_TGID;
372
373	/*
374	 * The taskstats structure is internally aligned on 8 byte
375	 * boundaries but the layout of the aggregrate reply, with
376	 * two NLA headers and the pid (each 4 bytes), actually
377	 * force the entire structure to be unaligned. This causes
378	 * the kernel to issue unaligned access warnings on some
379	 * architectures like ia64. Unfortunately, some software out there
380	 * doesn't properly unroll the NLA packet and assumes that the start
381	 * of the taskstats structure will always be 20 bytes from the start
382	 * of the netlink payload. Aligning the start of the taskstats
383	 * structure breaks this software, which we don't want. So, for now
384	 * the alignment only happens on architectures that require it
385	 * and those users will have to update to fixed versions of those
386	 * packages. Space is reserved in the packet only when needed.
387	 * This ifdef should be removed in several years e.g. 2012 once
388	 * we can be confident that fixed versions are installed on most
389	 * systems. We add the padding before the aggregate since the
390	 * aggregate is already a defined type.
391	 */
392#ifdef TASKSTATS_NEEDS_PADDING
393	if (nla_put(skb, TASKSTATS_TYPE_NULL, 0, NULL) < 0)
394		goto err;
395#endif
396	na = nla_nest_start(skb, aggr);
397	if (!na)
398		goto err;
399
400	if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
401		nla_nest_cancel(skb, na);
402		goto err;
403	}
404	ret = nla_reserve(skb, TASKSTATS_TYPE_STATS, sizeof(struct taskstats));
405	if (!ret) {
406		nla_nest_cancel(skb, na);
407		goto err;
408	}
409	nla_nest_end(skb, na);
410
411	return nla_data(ret);
412err:
413	return NULL;
414}
415
416static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
417{
418	int rc = 0;
419	struct sk_buff *rep_skb;
420	struct cgroupstats *stats;
421	struct nlattr *na;
422	size_t size;
423	u32 fd;
424	struct fd f;
425
426	na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
427	if (!na)
428		return -EINVAL;
429
430	fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
431	f = fdget(fd);
432	if (!f.file)
433		return 0;
434
435	size = nla_total_size(sizeof(struct cgroupstats));
436
437	rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
438				size);
439	if (rc < 0)
440		goto err;
441
442	na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
443				sizeof(struct cgroupstats));
444	if (na == NULL) {
445		nlmsg_free(rep_skb);
446		rc = -EMSGSIZE;
447		goto err;
448	}
449
450	stats = nla_data(na);
451	memset(stats, 0, sizeof(*stats));
452
453	rc = cgroupstats_build(stats, f.file->f_path.dentry);
454	if (rc < 0) {
455		nlmsg_free(rep_skb);
456		goto err;
457	}
458
459	rc = send_reply(rep_skb, info);
460
461err:
462	fdput(f);
463	return rc;
464}
465
466static int cmd_attr_register_cpumask(struct genl_info *info)
467{
468	cpumask_var_t mask;
469	int rc;
470
471	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
472		return -ENOMEM;
473	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
474	if (rc < 0)
475		goto out;
476	rc = add_del_listener(info->snd_portid, mask, REGISTER);
477out:
478	free_cpumask_var(mask);
479	return rc;
480}
481
482static int cmd_attr_deregister_cpumask(struct genl_info *info)
483{
484	cpumask_var_t mask;
485	int rc;
486
487	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
488		return -ENOMEM;
489	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
490	if (rc < 0)
491		goto out;
492	rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
493out:
494	free_cpumask_var(mask);
495	return rc;
496}
497
498static size_t taskstats_packet_size(void)
499{
500	size_t size;
501
502	size = nla_total_size(sizeof(u32)) +
503		nla_total_size(sizeof(struct taskstats)) + nla_total_size(0);
504#ifdef TASKSTATS_NEEDS_PADDING
505	size += nla_total_size(0); /* Padding for alignment */
506#endif
507	return size;
508}
509
510static int cmd_attr_pid(struct genl_info *info)
511{
512	struct taskstats *stats;
513	struct sk_buff *rep_skb;
514	size_t size;
515	u32 pid;
516	int rc;
517
518	size = taskstats_packet_size();
519
520	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
521	if (rc < 0)
522		return rc;
523
524	rc = -EINVAL;
525	pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
526	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
527	if (!stats)
528		goto err;
529
530	rc = fill_stats_for_pid(pid, stats);
531	if (rc < 0)
532		goto err;
533	return send_reply(rep_skb, info);
534err:
535	nlmsg_free(rep_skb);
536	return rc;
537}
538
539static int cmd_attr_tgid(struct genl_info *info)
540{
541	struct taskstats *stats;
542	struct sk_buff *rep_skb;
543	size_t size;
544	u32 tgid;
545	int rc;
546
547	size = taskstats_packet_size();
548
549	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
550	if (rc < 0)
551		return rc;
552
553	rc = -EINVAL;
554	tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
555	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
556	if (!stats)
557		goto err;
558
559	rc = fill_stats_for_tgid(tgid, stats);
560	if (rc < 0)
561		goto err;
562	return send_reply(rep_skb, info);
563err:
564	nlmsg_free(rep_skb);
565	return rc;
566}
567
568static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
569{
570	if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
571		return cmd_attr_register_cpumask(info);
572	else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
573		return cmd_attr_deregister_cpumask(info);
574	else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
575		return cmd_attr_pid(info);
576	else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
577		return cmd_attr_tgid(info);
578	else
579		return -EINVAL;
580}
581
582static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
583{
584	struct signal_struct *sig = tsk->signal;
585	struct taskstats *stats;
586
587	if (sig->stats || thread_group_empty(tsk))
588		goto ret;
589
590	/* No problem if kmem_cache_zalloc() fails */
591	stats = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
592
593	spin_lock_irq(&tsk->sighand->siglock);
594	if (!sig->stats) {
595		sig->stats = stats;
596		stats = NULL;
597	}
598	spin_unlock_irq(&tsk->sighand->siglock);
599
600	if (stats)
601		kmem_cache_free(taskstats_cache, stats);
602ret:
603	return sig->stats;
604}
605
606/* Send pid data out on exit */
607void taskstats_exit(struct task_struct *tsk, int group_dead)
608{
609	int rc;
610	struct listener_list *listeners;
611	struct taskstats *stats;
612	struct sk_buff *rep_skb;
613	size_t size;
614	int is_thread_group;
615
616	if (!family_registered)
617		return;
618
619	/*
620	 * Size includes space for nested attributes
621	 */
622	size = taskstats_packet_size();
623
624	is_thread_group = !!taskstats_tgid_alloc(tsk);
625	if (is_thread_group) {
626		/* PID + STATS + TGID + STATS */
627		size = 2 * size;
628		/* fill the tsk->signal->stats structure */
629		fill_tgid_exit(tsk);
630	}
631
632	listeners = raw_cpu_ptr(&listener_array);
633	if (list_empty(&listeners->list))
634		return;
635
636	rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
637	if (rc < 0)
638		return;
639
640	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
641			 task_pid_nr_ns(tsk, &init_pid_ns));
642	if (!stats)
643		goto err;
644
645	fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
646
647	/*
648	 * Doesn't matter if tsk is the leader or the last group member leaving
649	 */
650	if (!is_thread_group || !group_dead)
651		goto send;
652
653	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
654			 task_tgid_nr_ns(tsk, &init_pid_ns));
655	if (!stats)
656		goto err;
657
658	memcpy(stats, tsk->signal->stats, sizeof(*stats));
659
660send:
661	send_cpu_listeners(rep_skb, listeners);
662	return;
663err:
664	nlmsg_free(rep_skb);
665}
666
667static const struct genl_ops taskstats_ops[] = {
668	{
669		.cmd		= TASKSTATS_CMD_GET,
670		.doit		= taskstats_user_cmd,
671		.policy		= taskstats_cmd_get_policy,
672		.flags		= GENL_ADMIN_PERM,
673	},
674	{
675		.cmd		= CGROUPSTATS_CMD_GET,
676		.doit		= cgroupstats_user_cmd,
677		.policy		= cgroupstats_cmd_get_policy,
678	},
679};
680
681/* Needed early in initialization */
682void __init taskstats_init_early(void)
683{
684	unsigned int i;
685
686	taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
687	for_each_possible_cpu(i) {
688		INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
689		init_rwsem(&(per_cpu(listener_array, i).sem));
690	}
691}
692
693static int __init taskstats_init(void)
694{
695	int rc;
696
697	rc = genl_register_family_with_ops(&family, taskstats_ops);
698	if (rc)
699		return rc;
700
701	family_registered = 1;
702	pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
703	return 0;
704}
705
706/*
707 * late initcall ensures initialization of statistics collection
708 * mechanisms precedes initialization of the taskstats interface
709 */
710late_initcall(taskstats_init);
711