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	int ret = 0;
294
295	if (!cpumask_subset(mask, cpu_possible_mask))
296		return -EINVAL;
297
298	if (current_user_ns() != &init_user_ns)
299		return -EINVAL;
300
301	if (task_active_pid_ns(current) != &init_pid_ns)
302		return -EINVAL;
303
304	if (isadd == REGISTER) {
305		for_each_cpu(cpu, mask) {
306			s = kmalloc_node(sizeof(struct listener),
307					GFP_KERNEL, cpu_to_node(cpu));
308			if (!s) {
309				ret = -ENOMEM;
310				goto cleanup;
311			}
312			s->pid = pid;
313			s->valid = 1;
314
315			listeners = &per_cpu(listener_array, cpu);
316			down_write(&listeners->sem);
317			list_for_each_entry(s2, &listeners->list, list) {
318				if (s2->pid == pid && s2->valid)
319					goto exists;
320			}
321			list_add(&s->list, &listeners->list);
322			s = NULL;
323exists:
324			up_write(&listeners->sem);
325			kfree(s); /* nop if NULL */
326		}
327		return 0;
328	}
329
330	/* Deregister or cleanup */
331cleanup:
332	for_each_cpu(cpu, mask) {
333		listeners = &per_cpu(listener_array, cpu);
334		down_write(&listeners->sem);
335		list_for_each_entry_safe(s, tmp, &listeners->list, list) {
336			if (s->pid == pid) {
337				list_del(&s->list);
338				kfree(s);
339				break;
340			}
341		}
342		up_write(&listeners->sem);
343	}
344	return ret;
345}
346
347static int parse(struct nlattr *na, struct cpumask *mask)
348{
349	char *data;
350	int len;
351	int ret;
352
353	if (na == NULL)
354		return 1;
355	len = nla_len(na);
356	if (len > TASKSTATS_CPUMASK_MAXLEN)
357		return -E2BIG;
358	if (len < 1)
359		return -EINVAL;
360	data = kmalloc(len, GFP_KERNEL);
361	if (!data)
362		return -ENOMEM;
363	nla_strlcpy(data, na, len);
364	ret = cpulist_parse(data, mask);
365	kfree(data);
366	return ret;
367}
368
369#if defined(CONFIG_64BIT) && !defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
370#define TASKSTATS_NEEDS_PADDING 1
371#endif
372
373static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
374{
375	struct nlattr *na, *ret;
376	int aggr;
377
378	aggr = (type == TASKSTATS_TYPE_PID)
379			? TASKSTATS_TYPE_AGGR_PID
380			: TASKSTATS_TYPE_AGGR_TGID;
381
382	/*
383	 * The taskstats structure is internally aligned on 8 byte
384	 * boundaries but the layout of the aggregrate reply, with
385	 * two NLA headers and the pid (each 4 bytes), actually
386	 * force the entire structure to be unaligned. This causes
387	 * the kernel to issue unaligned access warnings on some
388	 * architectures like ia64. Unfortunately, some software out there
389	 * doesn't properly unroll the NLA packet and assumes that the start
390	 * of the taskstats structure will always be 20 bytes from the start
391	 * of the netlink payload. Aligning the start of the taskstats
392	 * structure breaks this software, which we don't want. So, for now
393	 * the alignment only happens on architectures that require it
394	 * and those users will have to update to fixed versions of those
395	 * packages. Space is reserved in the packet only when needed.
396	 * This ifdef should be removed in several years e.g. 2012 once
397	 * we can be confident that fixed versions are installed on most
398	 * systems. We add the padding before the aggregate since the
399	 * aggregate is already a defined type.
400	 */
401#ifdef TASKSTATS_NEEDS_PADDING
402	if (nla_put(skb, TASKSTATS_TYPE_NULL, 0, NULL) < 0)
403		goto err;
404#endif
405	na = nla_nest_start(skb, aggr);
406	if (!na)
407		goto err;
408
409	if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
410		nla_nest_cancel(skb, na);
411		goto err;
412	}
413	ret = nla_reserve(skb, TASKSTATS_TYPE_STATS, sizeof(struct taskstats));
414	if (!ret) {
415		nla_nest_cancel(skb, na);
416		goto err;
417	}
418	nla_nest_end(skb, na);
419
420	return nla_data(ret);
421err:
422	return NULL;
423}
424
425static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
426{
427	int rc = 0;
428	struct sk_buff *rep_skb;
429	struct cgroupstats *stats;
430	struct nlattr *na;
431	size_t size;
432	u32 fd;
433	struct fd f;
434
435	na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
436	if (!na)
437		return -EINVAL;
438
439	fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
440	f = fdget(fd);
441	if (!f.file)
442		return 0;
443
444	size = nla_total_size(sizeof(struct cgroupstats));
445
446	rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
447				size);
448	if (rc < 0)
449		goto err;
450
451	na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
452				sizeof(struct cgroupstats));
453	if (na == NULL) {
454		nlmsg_free(rep_skb);
455		rc = -EMSGSIZE;
456		goto err;
457	}
458
459	stats = nla_data(na);
460	memset(stats, 0, sizeof(*stats));
461
462	rc = cgroupstats_build(stats, f.file->f_dentry);
463	if (rc < 0) {
464		nlmsg_free(rep_skb);
465		goto err;
466	}
467
468	rc = send_reply(rep_skb, info);
469
470err:
471	fdput(f);
472	return rc;
473}
474
475static int cmd_attr_register_cpumask(struct genl_info *info)
476{
477	cpumask_var_t mask;
478	int rc;
479
480	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
481		return -ENOMEM;
482	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
483	if (rc < 0)
484		goto out;
485	rc = add_del_listener(info->snd_portid, mask, REGISTER);
486out:
487	free_cpumask_var(mask);
488	return rc;
489}
490
491static int cmd_attr_deregister_cpumask(struct genl_info *info)
492{
493	cpumask_var_t mask;
494	int rc;
495
496	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
497		return -ENOMEM;
498	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
499	if (rc < 0)
500		goto out;
501	rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
502out:
503	free_cpumask_var(mask);
504	return rc;
505}
506
507static size_t taskstats_packet_size(void)
508{
509	size_t size;
510
511	size = nla_total_size(sizeof(u32)) +
512		nla_total_size(sizeof(struct taskstats)) + nla_total_size(0);
513#ifdef TASKSTATS_NEEDS_PADDING
514	size += nla_total_size(0); /* Padding for alignment */
515#endif
516	return size;
517}
518
519static int cmd_attr_pid(struct genl_info *info)
520{
521	struct taskstats *stats;
522	struct sk_buff *rep_skb;
523	size_t size;
524	u32 pid;
525	int rc;
526
527	size = taskstats_packet_size();
528
529	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
530	if (rc < 0)
531		return rc;
532
533	rc = -EINVAL;
534	pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
535	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
536	if (!stats)
537		goto err;
538
539	rc = fill_stats_for_pid(pid, stats);
540	if (rc < 0)
541		goto err;
542	return send_reply(rep_skb, info);
543err:
544	nlmsg_free(rep_skb);
545	return rc;
546}
547
548static int cmd_attr_tgid(struct genl_info *info)
549{
550	struct taskstats *stats;
551	struct sk_buff *rep_skb;
552	size_t size;
553	u32 tgid;
554	int rc;
555
556	size = taskstats_packet_size();
557
558	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
559	if (rc < 0)
560		return rc;
561
562	rc = -EINVAL;
563	tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
564	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
565	if (!stats)
566		goto err;
567
568	rc = fill_stats_for_tgid(tgid, stats);
569	if (rc < 0)
570		goto err;
571	return send_reply(rep_skb, info);
572err:
573	nlmsg_free(rep_skb);
574	return rc;
575}
576
577static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
578{
579	if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
580		return cmd_attr_register_cpumask(info);
581	else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
582		return cmd_attr_deregister_cpumask(info);
583	else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
584		return cmd_attr_pid(info);
585	else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
586		return cmd_attr_tgid(info);
587	else
588		return -EINVAL;
589}
590
591static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
592{
593	struct signal_struct *sig = tsk->signal;
594	struct taskstats *stats;
595
596	if (sig->stats || thread_group_empty(tsk))
597		goto ret;
598
599	/* No problem if kmem_cache_zalloc() fails */
600	stats = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
601
602	spin_lock_irq(&tsk->sighand->siglock);
603	if (!sig->stats) {
604		sig->stats = stats;
605		stats = NULL;
606	}
607	spin_unlock_irq(&tsk->sighand->siglock);
608
609	if (stats)
610		kmem_cache_free(taskstats_cache, stats);
611ret:
612	return sig->stats;
613}
614
615/* Send pid data out on exit */
616void taskstats_exit(struct task_struct *tsk, int group_dead)
617{
618	int rc;
619	struct listener_list *listeners;
620	struct taskstats *stats;
621	struct sk_buff *rep_skb;
622	size_t size;
623	int is_thread_group;
624
625	if (!family_registered)
626		return;
627
628	/*
629	 * Size includes space for nested attributes
630	 */
631	size = taskstats_packet_size();
632
633	is_thread_group = !!taskstats_tgid_alloc(tsk);
634	if (is_thread_group) {
635		/* PID + STATS + TGID + STATS */
636		size = 2 * size;
637		/* fill the tsk->signal->stats structure */
638		fill_tgid_exit(tsk);
639	}
640
641	listeners = raw_cpu_ptr(&listener_array);
642	if (list_empty(&listeners->list))
643		return;
644
645	rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
646	if (rc < 0)
647		return;
648
649	stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
650			 task_pid_nr_ns(tsk, &init_pid_ns));
651	if (!stats)
652		goto err;
653
654	fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
655
656	/*
657	 * Doesn't matter if tsk is the leader or the last group member leaving
658	 */
659	if (!is_thread_group || !group_dead)
660		goto send;
661
662	stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
663			 task_tgid_nr_ns(tsk, &init_pid_ns));
664	if (!stats)
665		goto err;
666
667	memcpy(stats, tsk->signal->stats, sizeof(*stats));
668
669send:
670	send_cpu_listeners(rep_skb, listeners);
671	return;
672err:
673	nlmsg_free(rep_skb);
674}
675
676static const struct genl_ops taskstats_ops[] = {
677	{
678		.cmd		= TASKSTATS_CMD_GET,
679		.doit		= taskstats_user_cmd,
680		.policy		= taskstats_cmd_get_policy,
681		.flags		= GENL_ADMIN_PERM,
682	},
683	{
684		.cmd		= CGROUPSTATS_CMD_GET,
685		.doit		= cgroupstats_user_cmd,
686		.policy		= cgroupstats_cmd_get_policy,
687	},
688};
689
690/* Needed early in initialization */
691void __init taskstats_init_early(void)
692{
693	unsigned int i;
694
695	taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
696	for_each_possible_cpu(i) {
697		INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
698		init_rwsem(&(per_cpu(listener_array, i).sem));
699	}
700}
701
702static int __init taskstats_init(void)
703{
704	int rc;
705
706	rc = genl_register_family_with_ops(&family, taskstats_ops);
707	if (rc)
708		return rc;
709
710	family_registered = 1;
711	pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
712	return 0;
713}
714
715/*
716 * late initcall ensures initialization of statistics collection
717 * mechanisms precedes initialization of the taskstats interface
718 */
719late_initcall(taskstats_init);
720