projects/linux·blob v7.2-rc6 · refs

722 lines · 15.8 KiB

1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * taskstats.c - Export per-task statistics to userland
4 *
5 * Copyright (C) Shailabh Nagar, IBM Corp. 2006
6 * (C) Balbir Singh, IBM Corp. 2006
7 */
8
9#include <linux/kernel.h>
10#include <linux/taskstats_kern.h>
11#include <linux/tsacct_kern.h>
12#include <linux/acct.h>
13#include <linux/delayacct.h>
14#include <linux/cpumask.h>
15#include <linux/percpu.h>
16#include <linux/slab.h>
17#include <linux/cgroupstats.h>
18#include <linux/cgroup.h>
19#include <linux/fs.h>
20#include <linux/file.h>
21#include <linux/pid_namespace.h>
22#include <net/genetlink.h>
23#include <linux/atomic.h>
24#include <linux/sched/cputime.h>
25
26/*
27 * Maximum length of a cpumask that can be specified in
28 * the TASKSTATS_CMD_ATTR_REGISTER/DEREGISTER_CPUMASK attribute
29 */
30#define TASKSTATS_CPUMASK_MAXLEN (100+6*NR_CPUS)
31
32static DEFINE_PER_CPU(__u32, taskstats_seqnum);
33static int family_registered;
34struct kmem_cache *taskstats_cache;
35
36static struct genl_family family;
37
38static const struct nla_policy taskstats_cmd_get_policy[] = {
39 [TASKSTATS_CMD_ATTR_PID] = { .type = NLA_U32 },
40 [TASKSTATS_CMD_ATTR_TGID] = { .type = NLA_U32 },
41 [TASKSTATS_CMD_ATTR_REGISTER_CPUMASK] = { .type = NLA_STRING },
42 [TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK] = { .type = NLA_STRING },};
43
44static const struct nla_policy cgroupstats_cmd_get_policy[] = {
45 [CGROUPSTATS_CMD_ATTR_FD] = { .type = NLA_U32 },
46};
47
48struct listener {
49 struct list_head list;
50 pid_t pid;
51 char valid;
52};
53
54struct listener_list {
55 struct rw_semaphore sem;
56 struct list_head list;
57};
58static DEFINE_PER_CPU(struct listener_list, listener_array);
59
60enum actions {
61 REGISTER,
62 DEREGISTER,
63 CPU_DONT_CARE
64};
65
66static int prepare_reply(struct genl_info *info, u8 cmd, struct sk_buff **skbp,
67 size_t size)
68{
69 struct sk_buff *skb;
70 void *reply;
71
72 /*
73 * If new attributes are added, please revisit this allocation
74 */
75 skb = genlmsg_new(size, GFP_KERNEL);
76 if (!skb)
77 return -ENOMEM;
78
79 if (!info) {
80 int seq = this_cpu_inc_return(taskstats_seqnum) - 1;
81
82 reply = genlmsg_put(skb, 0, seq, &family, 0, cmd);
83 } else
84 reply = genlmsg_put_reply(skb, info, &family, 0, cmd);
85 if (reply == NULL) {
86 nlmsg_free(skb);
87 return -EINVAL;
88 }
89
90 *skbp = skb;
91 return 0;
92}
93
94/*
95 * Send taskstats data in @skb to listener with nl_pid @pid
96 */
97static int send_reply(struct sk_buff *skb, struct genl_info *info)
98{
99 struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
100 void *reply = genlmsg_data(genlhdr);
101
102 genlmsg_end(skb, reply);
103
104 return genlmsg_reply(skb, info);
105}
106
107/*
108 * Send taskstats data in @skb to listeners registered for @cpu's exit data
109 */
110static void send_cpu_listeners(struct sk_buff *skb,
111 struct listener_list *listeners)
112{
113 struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
114 struct listener *s, *tmp;
115 struct sk_buff *skb_next, *skb_cur = skb;
116 void *reply = genlmsg_data(genlhdr);
117 int delcount = 0;
118
119 genlmsg_end(skb, reply);
120
121 down_read(&listeners->sem);
122 list_for_each_entry(s, &listeners->list, list) {
123 int rc;
124
125 skb_next = NULL;
126 if (!list_is_last(&s->list, &listeners->list)) {
127 skb_next = skb_clone(skb_cur, GFP_KERNEL);
128 if (!skb_next)
129 break;
130 }
131 rc = genlmsg_unicast(&init_net, skb_cur, s->pid);
132 if (rc == -ECONNREFUSED) {
133 s->valid = 0;
134 delcount++;
135 }
136 skb_cur = skb_next;
137 }
138 up_read(&listeners->sem);
139
140 if (skb_cur)
141 nlmsg_free(skb_cur);
142
143 if (!delcount)
144 return;
145
146 /* Delete invalidated entries */
147 down_write(&listeners->sem);
148 list_for_each_entry_safe(s, tmp, &listeners->list, list) {
149 if (!s->valid) {
150 list_del(&s->list);
151 kfree(s);
152 }
153 }
154 up_write(&listeners->sem);
155}
156
157static void exe_add_tsk(struct taskstats *stats, struct task_struct *tsk)
158{
159 /* No idea if I'm allowed to access that here, now. */
160 struct file *exe_file = get_task_exe_file(tsk);
161
162 if (exe_file) {
163 /* Following cp_new_stat64() in stat.c . */
164 stats->ac_exe_dev =
165 huge_encode_dev(exe_file->f_inode->i_sb->s_dev);
166 stats->ac_exe_inode = exe_file->f_inode->i_ino;
167 fput(exe_file);
168 } else {
169 stats->ac_exe_dev = 0;
170 stats->ac_exe_inode = 0;
171 }
172}
173
174static void fill_stats(struct user_namespace *user_ns,
175 struct pid_namespace *pid_ns,
176 struct task_struct *tsk, struct taskstats *stats)
177{
178 memset(stats, 0, sizeof(*stats));
179 /*
180 * Each accounting subsystem adds calls to its functions to
181 * fill in relevant parts of struct taskstsats as follows
182 *
183 * per-task-foo(stats, tsk);
184 */
185
186 delayacct_add_tsk(stats, tsk);
187
188 /* fill in basic acct fields */
189 stats->version = TASKSTATS_VERSION;
190 stats->nvcsw = tsk->nvcsw;
191 stats->nivcsw = tsk->nivcsw;
192 bacct_add_tsk(user_ns, pid_ns, stats, tsk);
193
194 /* fill in extended acct fields */
195 xacct_add_tsk(stats, tsk);
196
197 /* add executable info */
198 exe_add_tsk(stats, tsk);
199}
200
201static int fill_stats_for_pid(pid_t pid, struct taskstats *stats)
202{
203 struct task_struct *tsk;
204
205 tsk = find_get_task_by_vpid(pid);
206 if (!tsk)
207 return -ESRCH;
208 fill_stats(current_user_ns(), task_active_pid_ns(current), tsk, stats);
209 put_task_struct(tsk);
210 return 0;
211}
212
213static void tgid_stats_add_task(struct taskstats *stats,
214 struct task_struct *tsk, u64 now_ns)
215{
216 u64 delta, utime, stime;
217
218 /*
219 * Each accounting subsystem calls its functions here to
220 * accumulate its per-task stats for tsk, into the per-tgid structure
221 *
222 * per-task-foo(stats, tsk);
223 */
224 delayacct_add_tsk(stats, tsk);
225
226 /* calculate task elapsed time in nsec */
227 delta = now_ns - tsk->start_time;
228 /* Convert to micro seconds */
229 do_div(delta, NSEC_PER_USEC);
230 stats->ac_etime += delta;
231
232 task_cputime(tsk, &utime, &stime);
233 stats->ac_utime += div_u64(utime, NSEC_PER_USEC);
234 stats->ac_stime += div_u64(stime, NSEC_PER_USEC);
235
236 stats->nvcsw += tsk->nvcsw;
237 stats->nivcsw += tsk->nivcsw;
238}
239
240static int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
241{
242 struct task_struct *tsk, *first;
243 unsigned long flags;
244 int rc = -ESRCH;
245 u64 now_ns;
246
247 /*
248 * Add additional stats from live tasks except zombie thread group
249 * leaders who are already counted with the dead tasks
250 */
251 rcu_read_lock();
252 first = find_task_by_vpid(tgid);
253
254 if (!first || !lock_task_sighand(first, &flags))
255 goto out;
256
257 if (first->signal->stats)
258 memcpy(stats, first->signal->stats, sizeof(*stats));
259 else
260 memset(stats, 0, sizeof(*stats));
261
262 now_ns = ktime_get_ns();
263 for_each_thread(first, tsk) {
264 if (tsk->exit_state)
265 continue;
266
267 tgid_stats_add_task(stats, tsk, now_ns);
268 }
269
270 unlock_task_sighand(first, &flags);
271 rc = 0;
272out:
273 rcu_read_unlock();
274
275 stats->version = TASKSTATS_VERSION;
276 /*
277 * Accounting subsystems can also add calls here to modify
278 * fields of taskstats.
279 */
280 return rc;
281}
282
283static void fill_tgid_exit(struct task_struct *tsk)
284{
285 unsigned long flags;
286 u64 now_ns;
287
288 spin_lock_irqsave(&tsk->sighand->siglock, flags);
289 if (!tsk->signal->stats)
290 goto ret;
291
292 now_ns = ktime_get_ns();
293 tgid_stats_add_task(tsk->signal->stats, tsk, now_ns);
294ret:
295 spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
296 return;
297}
298
299static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
300{
301 struct listener_list *listeners;
302 struct listener *s, *tmp, *s2;
303 unsigned int cpu;
304 int ret = 0;
305
306 if (!cpumask_subset(mask, cpu_possible_mask))
307 return -EINVAL;
308
309 if (current_user_ns() != &init_user_ns)
310 return -EINVAL;
311
312 if (task_active_pid_ns(current) != &init_pid_ns)
313 return -EINVAL;
314
315 if (isadd == REGISTER) {
316 for_each_cpu(cpu, mask) {
317 s = kmalloc_node(sizeof(struct listener),
318 GFP_KERNEL, cpu_to_node(cpu));
319 if (!s) {
320 ret = -ENOMEM;
321 goto cleanup;
322 }
323 s->pid = pid;
324 s->valid = 1;
325
326 listeners = &per_cpu(listener_array, cpu);
327 down_write(&listeners->sem);
328 list_for_each_entry(s2, &listeners->list, list) {
329 if (s2->pid == pid && s2->valid)
330 goto exists;
331 }
332 list_add(&s->list, &listeners->list);
333 s = NULL;
334exists:
335 up_write(&listeners->sem);
336 kfree(s); /* nop if NULL */
337 }
338 return 0;
339 }
340
341 /* Deregister or cleanup */
342cleanup:
343 for_each_cpu(cpu, mask) {
344 listeners = &per_cpu(listener_array, cpu);
345 down_write(&listeners->sem);
346 list_for_each_entry_safe(s, tmp, &listeners->list, list) {
347 if (s->pid == pid) {
348 list_del(&s->list);
349 kfree(s);
350 break;
351 }
352 }
353 up_write(&listeners->sem);
354 }
355 return ret;
356}
357
358static int parse(struct nlattr *na, struct cpumask *mask)
359{
360 char *data;
361 int len;
362 int ret;
363
364 if (na == NULL)
365 return 1;
366 len = nla_len(na);
367 if (len > TASKSTATS_CPUMASK_MAXLEN)
368 return -E2BIG;
369 if (len < 1)
370 return -EINVAL;
371 data = kmalloc(len, GFP_KERNEL);
372 if (!data)
373 return -ENOMEM;
374 nla_strscpy(data, na, len);
375 ret = cpulist_parse(data, mask);
376 kfree(data);
377 return ret;
378}
379
380static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
381{
382 struct nlattr *na, *ret;
383 int aggr;
384
385 aggr = (type == TASKSTATS_TYPE_PID)
386 ? TASKSTATS_TYPE_AGGR_PID
387 : TASKSTATS_TYPE_AGGR_TGID;
388
389 na = nla_nest_start_noflag(skb, aggr);
390 if (!na)
391 goto err;
392
393 if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
394 nla_nest_cancel(skb, na);
395 goto err;
396 }
397 ret = nla_reserve_64bit(skb, TASKSTATS_TYPE_STATS,
398 sizeof(struct taskstats), TASKSTATS_TYPE_NULL);
399 if (!ret) {
400 nla_nest_cancel(skb, na);
401 goto err;
402 }
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
419 na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
420 if (!na)
421 return -EINVAL;
422
423 fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
424 CLASS(fd, f)(fd);
425 if (fd_empty(f))
426 return 0;
427
428 size = nla_total_size(sizeof(struct cgroupstats));
429
430 rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
431 size);
432 if (rc < 0)
433 return rc;
434
435 na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
436 sizeof(struct cgroupstats));
437 if (na == NULL) {
438 nlmsg_free(rep_skb);
439 return -EMSGSIZE;
440 }
441
442 stats = nla_data(na);
443 memset(stats, 0, sizeof(*stats));
444
445 rc = cgroupstats_build(stats, fd_file(f)->f_path.dentry);
446 if (rc < 0) {
447 nlmsg_free(rep_skb);
448 return rc;
449 }
450
451 return send_reply(rep_skb, info);
452}
453
454static int cmd_attr_register_cpumask(struct genl_info *info)
455{
456 cpumask_var_t mask;
457 int rc;
458
459 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
460 return -ENOMEM;
461 rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
462 if (rc < 0)
463 goto out;
464 rc = add_del_listener(info->snd_portid, mask, REGISTER);
465out:
466 free_cpumask_var(mask);
467 return rc;
468}
469
470static int cmd_attr_deregister_cpumask(struct genl_info *info)
471{
472 cpumask_var_t mask;
473 int rc;
474
475 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
476 return -ENOMEM;
477 rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
478 if (rc < 0)
479 goto out;
480 rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
481out:
482 free_cpumask_var(mask);
483 return rc;
484}
485
486static size_t taskstats_packet_size(void)
487{
488 size_t size;
489
490 size = nla_total_size(sizeof(u32)) +
491 nla_total_size_64bit(sizeof(struct taskstats)) +
492 nla_total_size(0);
493
494 return size;
495}
496
497static int cmd_attr_pid(struct genl_info *info)
498{
499 struct taskstats *stats;
500 struct sk_buff *rep_skb;
501 size_t size;
502 u32 pid;
503 int rc;
504
505 size = taskstats_packet_size();
506
507 rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
508 if (rc < 0)
509 return rc;
510
511 rc = -EINVAL;
512 pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
513 stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
514 if (!stats)
515 goto err;
516
517 rc = fill_stats_for_pid(pid, stats);
518 if (rc < 0)
519 goto err;
520 return send_reply(rep_skb, info);
521err:
522 nlmsg_free(rep_skb);
523 return rc;
524}
525
526static int cmd_attr_tgid(struct genl_info *info)
527{
528 struct taskstats *stats;
529 struct sk_buff *rep_skb;
530 size_t size;
531 u32 tgid;
532 int rc;
533
534 size = taskstats_packet_size();
535
536 rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
537 if (rc < 0)
538 return rc;
539
540 rc = -EINVAL;
541 tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
542 stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
543 if (!stats)
544 goto err;
545
546 rc = fill_stats_for_tgid(tgid, stats);
547 if (rc < 0)
548 goto err;
549 return send_reply(rep_skb, info);
550err:
551 nlmsg_free(rep_skb);
552 return rc;
553}
554
555static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
556{
557 if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
558 return cmd_attr_register_cpumask(info);
559 else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
560 return cmd_attr_deregister_cpumask(info);
561 else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
562 return cmd_attr_pid(info);
563 else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
564 return cmd_attr_tgid(info);
565 else
566 return -EINVAL;
567}
568
569static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
570{
571 struct signal_struct *sig = tsk->signal;
572 struct taskstats *stats_new, *stats;
573
574 /* Pairs with smp_store_release() below. */
575 stats = smp_load_acquire(&sig->stats);
576 if (stats || thread_group_empty(tsk))
577 return stats;
578
579 /* No problem if kmem_cache_zalloc() fails */
580 stats_new = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
581
582 spin_lock_irq(&tsk->sighand->siglock);
583 stats = sig->stats;
584 if (!stats) {
585 /*
586 * Pairs with smp_store_release() above and order the
587 * kmem_cache_zalloc().
588 */
589 smp_store_release(&sig->stats, stats_new);
590 stats = stats_new;
591 stats_new = NULL;
592 }
593 spin_unlock_irq(&tsk->sighand->siglock);
594
595 if (stats_new)
596 kmem_cache_free(taskstats_cache, stats_new);
597
598 return stats;
599}
600
601/* Send pid data out on exit */
602void taskstats_exit(struct task_struct *tsk, int group_dead)
603{
604 int rc;
605 struct listener_list *listeners;
606 struct taskstats *stats;
607 struct sk_buff *rep_skb;
608 size_t size;
609 int is_thread_group;
610
611 if (!family_registered)
612 return;
613
614 /*
615 * Size includes space for nested attributes
616 */
617 size = taskstats_packet_size();
618
619 is_thread_group = !!taskstats_tgid_alloc(tsk);
620 if (is_thread_group) {
621 /* PID + STATS + TGID + STATS */
622 size = 2 * size;
623 /* fill the tsk->signal->stats structure */
624 fill_tgid_exit(tsk);
625 }
626
627 listeners = raw_cpu_ptr(&listener_array);
628 if (list_empty(&listeners->list))
629 return;
630
631 rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
632 if (rc < 0)
633 return;
634
635 stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
636 task_pid_nr_ns(tsk, &init_pid_ns));
637 if (!stats)
638 goto err;
639
640 fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
641 if (group_dead)
642 stats->ac_flag |= AGROUP;
643
644 /*
645 * Doesn't matter if tsk is the leader or the last group member leaving
646 */
647 if (!is_thread_group || !group_dead)
648 goto send;
649
650 stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
651 task_tgid_nr_ns(tsk, &init_pid_ns));
652 if (!stats)
653 goto err;
654
655 memcpy(stats, tsk->signal->stats, sizeof(*stats));
656 stats->version = TASKSTATS_VERSION;
657
658send:
659 send_cpu_listeners(rep_skb, listeners);
660 return;
661err:
662 nlmsg_free(rep_skb);
663}
664
665static const struct genl_ops taskstats_ops[] = {
666 {
667 .cmd = TASKSTATS_CMD_GET,
668 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
669 .doit = taskstats_user_cmd,
670 .policy = taskstats_cmd_get_policy,
671 .maxattr = ARRAY_SIZE(taskstats_cmd_get_policy) - 1,
672 .flags = GENL_ADMIN_PERM,
673 },
674 {
675 .cmd = CGROUPSTATS_CMD_GET,
676 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
677 .doit = cgroupstats_user_cmd,
678 .policy = cgroupstats_cmd_get_policy,
679 .maxattr = ARRAY_SIZE(cgroupstats_cmd_get_policy) - 1,
680 },
681};
682
683static struct genl_family family __ro_after_init = {
684 .name = TASKSTATS_GENL_NAME,
685 .version = TASKSTATS_GENL_VERSION,
686 .module = THIS_MODULE,
687 .ops = taskstats_ops,
688 .n_ops = ARRAY_SIZE(taskstats_ops),
689 .resv_start_op = CGROUPSTATS_CMD_GET + 1,
690 .netnsok = true,
691};
692
693/* Needed early in initialization */
694void __init taskstats_init_early(void)
695{
696 unsigned int i;
697
698 taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
699 for_each_possible_cpu(i) {
700 INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
701 init_rwsem(&(per_cpu(listener_array, i).sem));
702 }
703}
704
705static int __init taskstats_init(void)
706{
707 int rc;
708
709 rc = genl_register_family(&family);
710 if (rc)
711 return rc;
712
713 family_registered = 1;
714 pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
715 return 0;
716}
717
718/*
719 * late initcall ensures initialization of statistics collection
720 * mechanisms precedes initialization of the taskstats interface
721 */
722late_initcall(taskstats_init);