projects/linux·blob v6.2-rc8 · refs

724 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 int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
214{
215 struct task_struct *tsk, *first;
216 unsigned long flags;
217 int rc = -ESRCH;
218 u64 delta, utime, stime;
219 u64 start_time;
220
221 /*
222 * Add additional stats from live tasks except zombie thread group
223 * leaders who are already counted with the dead tasks
224 */
225 rcu_read_lock();
226 first = find_task_by_vpid(tgid);
227
228 if (!first || !lock_task_sighand(first, &flags))
229 goto out;
230
231 if (first->signal->stats)
232 memcpy(stats, first->signal->stats, sizeof(*stats));
233 else
234 memset(stats, 0, sizeof(*stats));
235
236 tsk = first;
237 start_time = ktime_get_ns();
238 do {
239 if (tsk->exit_state)
240 continue;
241 /*
242 * Accounting subsystem can call its functions here to
243 * fill in relevant parts of struct taskstsats as follows
244 *
245 * per-task-foo(stats, tsk);
246 */
247 delayacct_add_tsk(stats, tsk);
248
249 /* calculate task elapsed time in nsec */
250 delta = start_time - tsk->start_time;
251 /* Convert to micro seconds */
252 do_div(delta, NSEC_PER_USEC);
253 stats->ac_etime += delta;
254
255 task_cputime(tsk, &utime, &stime);
256 stats->ac_utime += div_u64(utime, NSEC_PER_USEC);
257 stats->ac_stime += div_u64(stime, NSEC_PER_USEC);
258
259 stats->nvcsw += tsk->nvcsw;
260 stats->nivcsw += tsk->nivcsw;
261 } while_each_thread(first, tsk);
262
263 unlock_task_sighand(first, &flags);
264 rc = 0;
265out:
266 rcu_read_unlock();
267
268 stats->version = TASKSTATS_VERSION;
269 /*
270 * Accounting subsystems can also add calls here to modify
271 * fields of taskstats.
272 */
273 return rc;
274}
275
276static void fill_tgid_exit(struct task_struct *tsk)
277{
278 unsigned long flags;
279
280 spin_lock_irqsave(&tsk->sighand->siglock, flags);
281 if (!tsk->signal->stats)
282 goto ret;
283
284 /*
285 * Each accounting subsystem calls its functions here to
286 * accumalate its per-task stats for tsk, into the per-tgid structure
287 *
288 * per-task-foo(tsk->signal->stats, tsk);
289 */
290 delayacct_add_tsk(tsk->signal->stats, tsk);
291ret:
292 spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
293 return;
294}
295
296static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
297{
298 struct listener_list *listeners;
299 struct listener *s, *tmp, *s2;
300 unsigned int cpu;
301 int ret = 0;
302
303 if (!cpumask_subset(mask, cpu_possible_mask))
304 return -EINVAL;
305
306 if (current_user_ns() != &init_user_ns)
307 return -EINVAL;
308
309 if (task_active_pid_ns(current) != &init_pid_ns)
310 return -EINVAL;
311
312 if (isadd == REGISTER) {
313 for_each_cpu(cpu, mask) {
314 s = kmalloc_node(sizeof(struct listener),
315 GFP_KERNEL, cpu_to_node(cpu));
316 if (!s) {
317 ret = -ENOMEM;
318 goto cleanup;
319 }
320 s->pid = pid;
321 s->valid = 1;
322
323 listeners = &per_cpu(listener_array, cpu);
324 down_write(&listeners->sem);
325 list_for_each_entry(s2, &listeners->list, list) {
326 if (s2->pid == pid && s2->valid)
327 goto exists;
328 }
329 list_add(&s->list, &listeners->list);
330 s = NULL;
331exists:
332 up_write(&listeners->sem);
333 kfree(s); /* nop if NULL */
334 }
335 return 0;
336 }
337
338 /* Deregister or cleanup */
339cleanup:
340 for_each_cpu(cpu, mask) {
341 listeners = &per_cpu(listener_array, cpu);
342 down_write(&listeners->sem);
343 list_for_each_entry_safe(s, tmp, &listeners->list, list) {
344 if (s->pid == pid) {
345 list_del(&s->list);
346 kfree(s);
347 break;
348 }
349 }
350 up_write(&listeners->sem);
351 }
352 return ret;
353}
354
355static int parse(struct nlattr *na, struct cpumask *mask)
356{
357 char *data;
358 int len;
359 int ret;
360
361 if (na == NULL)
362 return 1;
363 len = nla_len(na);
364 if (len > TASKSTATS_CPUMASK_MAXLEN)
365 return -E2BIG;
366 if (len < 1)
367 return -EINVAL;
368 data = kmalloc(len, GFP_KERNEL);
369 if (!data)
370 return -ENOMEM;
371 nla_strscpy(data, na, len);
372 ret = cpulist_parse(data, mask);
373 kfree(data);
374 return ret;
375}
376
377static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
378{
379 struct nlattr *na, *ret;
380 int aggr;
381
382 aggr = (type == TASKSTATS_TYPE_PID)
383 ? TASKSTATS_TYPE_AGGR_PID
384 : TASKSTATS_TYPE_AGGR_TGID;
385
386 na = nla_nest_start_noflag(skb, aggr);
387 if (!na)
388 goto err;
389
390 if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
391 nla_nest_cancel(skb, na);
392 goto err;
393 }
394 ret = nla_reserve_64bit(skb, TASKSTATS_TYPE_STATS,
395 sizeof(struct taskstats), TASKSTATS_TYPE_NULL);
396 if (!ret) {
397 nla_nest_cancel(skb, na);
398 goto err;
399 }
400 nla_nest_end(skb, na);
401
402 return nla_data(ret);
403err:
404 return NULL;
405}
406
407static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
408{
409 int rc = 0;
410 struct sk_buff *rep_skb;
411 struct cgroupstats *stats;
412 struct nlattr *na;
413 size_t size;
414 u32 fd;
415 struct fd f;
416
417 na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
418 if (!na)
419 return -EINVAL;
420
421 fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
422 f = fdget(fd);
423 if (!f.file)
424 return 0;
425
426 size = nla_total_size(sizeof(struct cgroupstats));
427
428 rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
429 size);
430 if (rc < 0)
431 goto err;
432
433 na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
434 sizeof(struct cgroupstats));
435 if (na == NULL) {
436 nlmsg_free(rep_skb);
437 rc = -EMSGSIZE;
438 goto err;
439 }
440
441 stats = nla_data(na);
442 memset(stats, 0, sizeof(*stats));
443
444 rc = cgroupstats_build(stats, f.file->f_path.dentry);
445 if (rc < 0) {
446 nlmsg_free(rep_skb);
447 goto err;
448 }
449
450 rc = send_reply(rep_skb, info);
451
452err:
453 fdput(f);
454 return rc;
455}
456
457static int cmd_attr_register_cpumask(struct genl_info *info)
458{
459 cpumask_var_t mask;
460 int rc;
461
462 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
463 return -ENOMEM;
464 rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
465 if (rc < 0)
466 goto out;
467 rc = add_del_listener(info->snd_portid, mask, REGISTER);
468out:
469 free_cpumask_var(mask);
470 return rc;
471}
472
473static int cmd_attr_deregister_cpumask(struct genl_info *info)
474{
475 cpumask_var_t mask;
476 int rc;
477
478 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
479 return -ENOMEM;
480 rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
481 if (rc < 0)
482 goto out;
483 rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
484out:
485 free_cpumask_var(mask);
486 return rc;
487}
488
489static size_t taskstats_packet_size(void)
490{
491 size_t size;
492
493 size = nla_total_size(sizeof(u32)) +
494 nla_total_size_64bit(sizeof(struct taskstats)) +
495 nla_total_size(0);
496
497 return size;
498}
499
500static int cmd_attr_pid(struct genl_info *info)
501{
502 struct taskstats *stats;
503 struct sk_buff *rep_skb;
504 size_t size;
505 u32 pid;
506 int rc;
507
508 size = taskstats_packet_size();
509
510 rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
511 if (rc < 0)
512 return rc;
513
514 rc = -EINVAL;
515 pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
516 stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
517 if (!stats)
518 goto err;
519
520 rc = fill_stats_for_pid(pid, stats);
521 if (rc < 0)
522 goto err;
523 return send_reply(rep_skb, info);
524err:
525 nlmsg_free(rep_skb);
526 return rc;
527}
528
529static int cmd_attr_tgid(struct genl_info *info)
530{
531 struct taskstats *stats;
532 struct sk_buff *rep_skb;
533 size_t size;
534 u32 tgid;
535 int rc;
536
537 size = taskstats_packet_size();
538
539 rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
540 if (rc < 0)
541 return rc;
542
543 rc = -EINVAL;
544 tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
545 stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
546 if (!stats)
547 goto err;
548
549 rc = fill_stats_for_tgid(tgid, stats);
550 if (rc < 0)
551 goto err;
552 return send_reply(rep_skb, info);
553err:
554 nlmsg_free(rep_skb);
555 return rc;
556}
557
558static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
559{
560 if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
561 return cmd_attr_register_cpumask(info);
562 else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
563 return cmd_attr_deregister_cpumask(info);
564 else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
565 return cmd_attr_pid(info);
566 else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
567 return cmd_attr_tgid(info);
568 else
569 return -EINVAL;
570}
571
572static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
573{
574 struct signal_struct *sig = tsk->signal;
575 struct taskstats *stats_new, *stats;
576
577 /* Pairs with smp_store_release() below. */
578 stats = smp_load_acquire(&sig->stats);
579 if (stats || thread_group_empty(tsk))
580 return stats;
581
582 /* No problem if kmem_cache_zalloc() fails */
583 stats_new = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
584
585 spin_lock_irq(&tsk->sighand->siglock);
586 stats = sig->stats;
587 if (!stats) {
588 /*
589 * Pairs with smp_store_release() above and order the
590 * kmem_cache_zalloc().
591 */
592 smp_store_release(&sig->stats, stats_new);
593 stats = stats_new;
594 stats_new = NULL;
595 }
596 spin_unlock_irq(&tsk->sighand->siglock);
597
598 if (stats_new)
599 kmem_cache_free(taskstats_cache, stats_new);
600
601 return stats;
602}
603
604/* Send pid data out on exit */
605void taskstats_exit(struct task_struct *tsk, int group_dead)
606{
607 int rc;
608 struct listener_list *listeners;
609 struct taskstats *stats;
610 struct sk_buff *rep_skb;
611 size_t size;
612 int is_thread_group;
613
614 if (!family_registered)
615 return;
616
617 /*
618 * Size includes space for nested attributes
619 */
620 size = taskstats_packet_size();
621
622 is_thread_group = !!taskstats_tgid_alloc(tsk);
623 if (is_thread_group) {
624 /* PID + STATS + TGID + STATS */
625 size = 2 * size;
626 /* fill the tsk->signal->stats structure */
627 fill_tgid_exit(tsk);
628 }
629
630 listeners = raw_cpu_ptr(&listener_array);
631 if (list_empty(&listeners->list))
632 return;
633
634 rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
635 if (rc < 0)
636 return;
637
638 stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
639 task_pid_nr_ns(tsk, &init_pid_ns));
640 if (!stats)
641 goto err;
642
643 fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
644 if (group_dead)
645 stats->ac_flag |= AGROUP;
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 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
671 .doit = taskstats_user_cmd,
672 .policy = taskstats_cmd_get_policy,
673 .maxattr = ARRAY_SIZE(taskstats_cmd_get_policy) - 1,
674 .flags = GENL_ADMIN_PERM,
675 },
676 {
677 .cmd = CGROUPSTATS_CMD_GET,
678 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
679 .doit = cgroupstats_user_cmd,
680 .policy = cgroupstats_cmd_get_policy,
681 .maxattr = ARRAY_SIZE(cgroupstats_cmd_get_policy) - 1,
682 },
683};
684
685static struct genl_family family __ro_after_init = {
686 .name = TASKSTATS_GENL_NAME,
687 .version = TASKSTATS_GENL_VERSION,
688 .module = THIS_MODULE,
689 .ops = taskstats_ops,
690 .n_ops = ARRAY_SIZE(taskstats_ops),
691 .resv_start_op = CGROUPSTATS_CMD_GET + 1,
692 .netnsok = true,
693};
694
695/* Needed early in initialization */
696void __init taskstats_init_early(void)
697{
698 unsigned int i;
699
700 taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
701 for_each_possible_cpu(i) {
702 INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
703 init_rwsem(&(per_cpu(listener_array, i).sem));
704 }
705}
706
707static int __init taskstats_init(void)
708{
709 int rc;
710
711 rc = genl_register_family(&family);
712 if (rc)
713 return rc;
714
715 family_registered = 1;
716 pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
717 return 0;
718}
719
720/*
721 * late initcall ensures initialization of statistics collection
722 * mechanisms precedes initialization of the taskstats interface
723 */
724late_initcall(taskstats_init);