taskstats.c revision 9acc1853519a0473620d424105f9d49ea5b4e62e 
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/tsacct_kern.h>
24#include <linux/cpumask.h>
25#include <linux/percpu.h>
26#include <net/genetlink.h>
27#include <asm/atomic.h>
28
29/*
30 * Maximum length of a cpumask that can be specified in
31 * the TASKSTATS_CMD_ATTR_REGISTER/DEREGISTER_CPUMASK attribute
32 */
33#define TASKSTATS_CPUMASK_MAXLEN	(100+6*NR_CPUS)
34
35static DEFINE_PER_CPU(__u32, taskstats_seqnum) = { 0 };
36static int family_registered;
37kmem_cache_t *taskstats_cache;
38
39static struct genl_family family = {
40	.id		= GENL_ID_GENERATE,
41	.name		= TASKSTATS_GENL_NAME,
42	.version	= TASKSTATS_GENL_VERSION,
43	.maxattr	= TASKSTATS_CMD_ATTR_MAX,
44};
45
46static struct nla_policy taskstats_cmd_get_policy[TASKSTATS_CMD_ATTR_MAX+1]
47__read_mostly = {
48	[TASKSTATS_CMD_ATTR_PID]  = { .type = NLA_U32 },
49	[TASKSTATS_CMD_ATTR_TGID] = { .type = NLA_U32 },
50	[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK] = { .type = NLA_STRING },
51	[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK] = { .type = NLA_STRING },};
52
53struct listener {
54	struct list_head list;
55	pid_t pid;
56	char valid;
57};
58
59struct listener_list {
60	struct rw_semaphore sem;
61	struct list_head list;
62};
63static DEFINE_PER_CPU(struct listener_list, listener_array);
64
65enum actions {
66	REGISTER,
67	DEREGISTER,
68	CPU_DONT_CARE
69};
70
71static int prepare_reply(struct genl_info *info, u8 cmd, struct sk_buff **skbp,
72			void **replyp, size_t size)
73{
74	struct sk_buff *skb;
75	void *reply;
76
77	/*
78	 * If new attributes are added, please revisit this allocation
79	 */
80	skb = nlmsg_new(genlmsg_total_size(size), GFP_KERNEL);
81	if (!skb)
82		return -ENOMEM;
83
84	if (!info) {
85		int seq = get_cpu_var(taskstats_seqnum)++;
86		put_cpu_var(taskstats_seqnum);
87
88		reply = genlmsg_put(skb, 0, seq,
89				family.id, 0, 0,
90				cmd, family.version);
91	} else
92		reply = genlmsg_put(skb, info->snd_pid, info->snd_seq,
93				family.id, 0, 0,
94				cmd, family.version);
95	if (reply == NULL) {
96		nlmsg_free(skb);
97		return -EINVAL;
98	}
99
100	*skbp = skb;
101	*replyp = reply;
102	return 0;
103}
104
105/*
106 * Send taskstats data in @skb to listener with nl_pid @pid
107 */
108static int send_reply(struct sk_buff *skb, pid_t pid)
109{
110	struct genlmsghdr *genlhdr = nlmsg_data((struct nlmsghdr *)skb->data);
111	void *reply = genlmsg_data(genlhdr);
112	int rc;
113
114	rc = genlmsg_end(skb, reply);
115	if (rc < 0) {
116		nlmsg_free(skb);
117		return rc;
118	}
119
120	return genlmsg_unicast(skb, pid);
121}
122
123/*
124 * Send taskstats data in @skb to listeners registered for @cpu's exit data
125 */
126static void send_cpu_listeners(struct sk_buff *skb, unsigned int cpu)
127{
128	struct genlmsghdr *genlhdr = nlmsg_data((struct nlmsghdr *)skb->data);
129	struct listener_list *listeners;
130	struct listener *s, *tmp;
131	struct sk_buff *skb_next, *skb_cur = skb;
132	void *reply = genlmsg_data(genlhdr);
133	int rc, delcount = 0;
134
135	rc = genlmsg_end(skb, reply);
136	if (rc < 0) {
137		nlmsg_free(skb);
138		return;
139	}
140
141	rc = 0;
142	listeners = &per_cpu(listener_array, cpu);
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(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 int fill_pid(pid_t pid, struct task_struct *pidtsk,
178		struct taskstats *stats)
179{
180	int rc = 0;
181	struct task_struct *tsk = pidtsk;
182
183	if (!pidtsk) {
184		read_lock(&tasklist_lock);
185		tsk = find_task_by_pid(pid);
186		if (!tsk) {
187			read_unlock(&tasklist_lock);
188			return -ESRCH;
189		}
190		get_task_struct(tsk);
191		read_unlock(&tasklist_lock);
192	} else
193		get_task_struct(tsk);
194
195	/*
196	 * Each accounting subsystem adds calls to its functions to
197	 * fill in relevant parts of struct taskstsats as follows
198	 *
199	 *	per-task-foo(stats, tsk);
200	 */
201
202	delayacct_add_tsk(stats, tsk);
203
204	/* fill in basic acct fields */
205	stats->version = TASKSTATS_VERSION;
206	bacct_add_tsk(stats, tsk);
207
208	/* fill in extended acct fields */
209	xacct_add_tsk(stats, tsk);
210
211	/* Define err: label here if needed */
212	put_task_struct(tsk);
213	return rc;
214
215}
216
217static int fill_tgid(pid_t tgid, struct task_struct *tgidtsk,
218		struct taskstats *stats)
219{
220	struct task_struct *tsk, *first;
221	unsigned long flags;
222
223	/*
224	 * Add additional stats from live tasks except zombie thread group
225	 * leaders who are already counted with the dead tasks
226	 */
227	first = tgidtsk;
228	if (!first) {
229		read_lock(&tasklist_lock);
230		first = find_task_by_pid(tgid);
231		if (!first) {
232			read_unlock(&tasklist_lock);
233			return -ESRCH;
234		}
235		get_task_struct(first);
236		read_unlock(&tasklist_lock);
237	} else
238		get_task_struct(first);
239
240	/* Start with stats from dead tasks */
241	spin_lock_irqsave(&first->signal->stats_lock, flags);
242	if (first->signal->stats)
243		memcpy(stats, first->signal->stats, sizeof(*stats));
244	spin_unlock_irqrestore(&first->signal->stats_lock, flags);
245
246	tsk = first;
247	read_lock(&tasklist_lock);
248	do {
249		if (tsk->exit_state == EXIT_ZOMBIE && thread_group_leader(tsk))
250			continue;
251		/*
252		 * Accounting subsystem can call its functions here to
253		 * fill in relevant parts of struct taskstsats as follows
254		 *
255		 *	per-task-foo(stats, tsk);
256		 */
257		delayacct_add_tsk(stats, tsk);
258
259	} while_each_thread(first, tsk);
260	read_unlock(&tasklist_lock);
261	stats->version = TASKSTATS_VERSION;
262
263	/*
264	 * Accounting subsytems can also add calls here to modify
265	 * fields of taskstats.
266	 */
267
268	return 0;
269}
270
271
272static void fill_tgid_exit(struct task_struct *tsk)
273{
274	unsigned long flags;
275
276	spin_lock_irqsave(&tsk->signal->stats_lock, flags);
277	if (!tsk->signal->stats)
278		goto ret;
279
280	/*
281	 * Each accounting subsystem calls its functions here to
282	 * accumalate its per-task stats for tsk, into the per-tgid structure
283	 *
284	 *	per-task-foo(tsk->signal->stats, tsk);
285	 */
286	delayacct_add_tsk(tsk->signal->stats, tsk);
287ret:
288	spin_unlock_irqrestore(&tsk->signal->stats_lock, flags);
289	return;
290}
291
292static int add_del_listener(pid_t pid, cpumask_t *maskp, int isadd)
293{
294	struct listener_list *listeners;
295	struct listener *s, *tmp;
296	unsigned int cpu;
297	cpumask_t mask = *maskp;
298
299	if (!cpus_subset(mask, cpu_possible_map))
300		return -EINVAL;
301
302	if (isadd == REGISTER) {
303		for_each_cpu_mask(cpu, mask) {
304			s = kmalloc_node(sizeof(struct listener), GFP_KERNEL,
305					 cpu_to_node(cpu));
306			if (!s)
307				goto cleanup;
308			s->pid = pid;
309			INIT_LIST_HEAD(&s->list);
310			s->valid = 1;
311
312			listeners = &per_cpu(listener_array, cpu);
313			down_write(&listeners->sem);
314			list_add(&s->list, &listeners->list);
315			up_write(&listeners->sem);
316		}
317		return 0;
318	}
319
320	/* Deregister or cleanup */
321cleanup:
322	for_each_cpu_mask(cpu, mask) {
323		listeners = &per_cpu(listener_array, cpu);
324		down_write(&listeners->sem);
325		list_for_each_entry_safe(s, tmp, &listeners->list, list) {
326			if (s->pid == pid) {
327				list_del(&s->list);
328				kfree(s);
329				break;
330			}
331		}
332		up_write(&listeners->sem);
333	}
334	return 0;
335}
336
337static int parse(struct nlattr *na, cpumask_t *mask)
338{
339	char *data;
340	int len;
341	int ret;
342
343	if (na == NULL)
344		return 1;
345	len = nla_len(na);
346	if (len > TASKSTATS_CPUMASK_MAXLEN)
347		return -E2BIG;
348	if (len < 1)
349		return -EINVAL;
350	data = kmalloc(len, GFP_KERNEL);
351	if (!data)
352		return -ENOMEM;
353	nla_strlcpy(data, na, len);
354	ret = cpulist_parse(data, *mask);
355	kfree(data);
356	return ret;
357}
358
359static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
360{
361	int rc = 0;
362	struct sk_buff *rep_skb;
363	struct taskstats stats;
364	void *reply;
365	size_t size;
366	struct nlattr *na;
367	cpumask_t mask;
368
369	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], &mask);
370	if (rc < 0)
371		return rc;
372	if (rc == 0)
373		return add_del_listener(info->snd_pid, &mask, REGISTER);
374
375	rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], &mask);
376	if (rc < 0)
377		return rc;
378	if (rc == 0)
379		return add_del_listener(info->snd_pid, &mask, DEREGISTER);
380
381	/*
382	 * Size includes space for nested attributes
383	 */
384	size = nla_total_size(sizeof(u32)) +
385		nla_total_size(sizeof(struct taskstats)) + nla_total_size(0);
386
387	memset(&stats, 0, sizeof(stats));
388	rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, &reply, size);
389	if (rc < 0)
390		return rc;
391
392	if (info->attrs[TASKSTATS_CMD_ATTR_PID]) {
393		u32 pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
394		rc = fill_pid(pid, NULL, &stats);
395		if (rc < 0)
396			goto err;
397
398		na = nla_nest_start(rep_skb, TASKSTATS_TYPE_AGGR_PID);
399		NLA_PUT_U32(rep_skb, TASKSTATS_TYPE_PID, pid);
400		NLA_PUT_TYPE(rep_skb, struct taskstats, TASKSTATS_TYPE_STATS,
401				stats);
402	} else if (info->attrs[TASKSTATS_CMD_ATTR_TGID]) {
403		u32 tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
404		rc = fill_tgid(tgid, NULL, &stats);
405		if (rc < 0)
406			goto err;
407
408		na = nla_nest_start(rep_skb, TASKSTATS_TYPE_AGGR_TGID);
409		NLA_PUT_U32(rep_skb, TASKSTATS_TYPE_TGID, tgid);
410		NLA_PUT_TYPE(rep_skb, struct taskstats, TASKSTATS_TYPE_STATS,
411				stats);
412	} else {
413		rc = -EINVAL;
414		goto err;
415	}
416
417	nla_nest_end(rep_skb, na);
418
419	return send_reply(rep_skb, info->snd_pid);
420
421nla_put_failure:
422	return genlmsg_cancel(rep_skb, reply);
423err:
424	nlmsg_free(rep_skb);
425	return rc;
426}
427
428void taskstats_exit_alloc(struct taskstats **ptidstats, unsigned int *mycpu)
429{
430	struct listener_list *listeners;
431	struct taskstats *tmp;
432	/*
433	 * This is the cpu on which the task is exiting currently and will
434	 * be the one for which the exit event is sent, even if the cpu
435	 * on which this function is running changes later.
436	 */
437	*mycpu = raw_smp_processor_id();
438
439	*ptidstats = NULL;
440	tmp = kmem_cache_zalloc(taskstats_cache, SLAB_KERNEL);
441	if (!tmp)
442		return;
443
444	listeners = &per_cpu(listener_array, *mycpu);
445	down_read(&listeners->sem);
446	if (!list_empty(&listeners->list)) {
447		*ptidstats = tmp;
448		tmp = NULL;
449	}
450	up_read(&listeners->sem);
451	kfree(tmp);
452}
453
454/* Send pid data out on exit */
455void taskstats_exit_send(struct task_struct *tsk, struct taskstats *tidstats,
456			int group_dead, unsigned int mycpu)
457{
458	int rc;
459	struct sk_buff *rep_skb;
460	void *reply;
461	size_t size;
462	int is_thread_group;
463	struct nlattr *na;
464	unsigned long flags;
465
466	if (!family_registered || !tidstats)
467		return;
468
469	spin_lock_irqsave(&tsk->signal->stats_lock, flags);
470	is_thread_group = tsk->signal->stats ? 1 : 0;
471	spin_unlock_irqrestore(&tsk->signal->stats_lock, flags);
472
473	rc = 0;
474	/*
475	 * Size includes space for nested attributes
476	 */
477	size = nla_total_size(sizeof(u32)) +
478		nla_total_size(sizeof(struct taskstats)) + nla_total_size(0);
479
480	if (is_thread_group)
481		size = 2 * size;	/* PID + STATS + TGID + STATS */
482
483	rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, &reply, size);
484	if (rc < 0)
485		goto ret;
486
487	rc = fill_pid(tsk->pid, tsk, tidstats);
488	if (rc < 0)
489		goto err_skb;
490
491	na = nla_nest_start(rep_skb, TASKSTATS_TYPE_AGGR_PID);
492	NLA_PUT_U32(rep_skb, TASKSTATS_TYPE_PID, (u32)tsk->pid);
493	NLA_PUT_TYPE(rep_skb, struct taskstats, TASKSTATS_TYPE_STATS,
494			*tidstats);
495	nla_nest_end(rep_skb, na);
496
497	if (!is_thread_group)
498		goto send;
499
500	/*
501	 * tsk has/had a thread group so fill the tsk->signal->stats structure
502	 * Doesn't matter if tsk is the leader or the last group member leaving
503	 */
504
505	fill_tgid_exit(tsk);
506	if (!group_dead)
507		goto send;
508
509	na = nla_nest_start(rep_skb, TASKSTATS_TYPE_AGGR_TGID);
510	NLA_PUT_U32(rep_skb, TASKSTATS_TYPE_TGID, (u32)tsk->tgid);
511	/* No locking needed for tsk->signal->stats since group is dead */
512	NLA_PUT_TYPE(rep_skb, struct taskstats, TASKSTATS_TYPE_STATS,
513			*tsk->signal->stats);
514	nla_nest_end(rep_skb, na);
515
516send:
517	send_cpu_listeners(rep_skb, mycpu);
518	return;
519
520nla_put_failure:
521	genlmsg_cancel(rep_skb, reply);
522	goto ret;
523err_skb:
524	nlmsg_free(rep_skb);
525ret:
526	return;
527}
528
529static struct genl_ops taskstats_ops = {
530	.cmd		= TASKSTATS_CMD_GET,
531	.doit		= taskstats_user_cmd,
532	.policy		= taskstats_cmd_get_policy,
533};
534
535/* Needed early in initialization */
536void __init taskstats_init_early(void)
537{
538	unsigned int i;
539
540	taskstats_cache = kmem_cache_create("taskstats_cache",
541						sizeof(struct taskstats),
542						0, SLAB_PANIC, NULL, NULL);
543	for_each_possible_cpu(i) {
544		INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
545		init_rwsem(&(per_cpu(listener_array, i).sem));
546	}
547}
548
549static int __init taskstats_init(void)
550{
551	int rc;
552
553	rc = genl_register_family(&family);
554	if (rc)
555		return rc;
556
557	rc = genl_register_ops(&family, &taskstats_ops);
558	if (rc < 0)
559		goto err;
560
561	family_registered = 1;
562	return 0;
563err:
564	genl_unregister_family(&family);
565	return rc;
566}
567
568/*
569 * late initcall ensures initialization of statistics collection
570 * mechanisms precedes initialization of the taskstats interface
571 */
572late_initcall(taskstats_init);
573