projects/linux·blob master · refs

705 lines · 15.5 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 len = nla_len(na);
365 if (len > TASKSTATS_CPUMASK_MAXLEN)
366 return -E2BIG;
367 if (len < 1)
368 return -EINVAL;
369 data = nla_strdup(na, GFP_KERNEL);
370 if (!data)
371 return -ENOMEM;
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
416 na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
417 if (!na)
418 return -EINVAL;
419
420 fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
421 CLASS(fd, f)(fd);
422 if (fd_empty(f))
423 return -EBADF;
424
425 size = nla_total_size(sizeof(struct cgroupstats));
426
427 rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
428 size);
429 if (rc < 0)
430 return rc;
431
432 na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
433 sizeof(struct cgroupstats));
434 if (na == NULL) {
435 nlmsg_free(rep_skb);
436 return -EMSGSIZE;
437 }
438
439 stats = nla_data(na);
440 memset(stats, 0, sizeof(*stats));
441
442 rc = cgroupstats_build(stats, fd_file(f)->f_path.dentry);
443 if (rc < 0) {
444 nlmsg_free(rep_skb);
445 return rc;
446 }
447
448 return send_reply(rep_skb, info);
449}
450
451static int cmd_attr_cpumask(struct genl_info *info, int attr,
452 enum actions action)
453{
454 cpumask_var_t mask __free(free_cpumask_var) = CPUMASK_VAR_NULL;
455 int rc;
456
457 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
458 return -ENOMEM;
459 rc = parse(info->attrs[attr], mask);
460 if (rc < 0)
461 return rc;
462 return add_del_listener(info->snd_portid, mask, action);
463}
464
465static size_t taskstats_packet_size(void)
466{
467 size_t size;
468
469 size = nla_total_size(sizeof(u32)) +
470 nla_total_size_64bit(sizeof(struct taskstats)) +
471 nla_total_size(0);
472
473 return size;
474}
475
476static int cmd_attr_pid(struct genl_info *info)
477{
478 struct taskstats *stats;
479 struct sk_buff *rep_skb;
480 size_t size;
481 u32 pid;
482 int rc;
483
484 size = taskstats_packet_size();
485
486 rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
487 if (rc < 0)
488 return rc;
489
490 rc = -EINVAL;
491 pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
492 stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
493 if (!stats)
494 goto err;
495
496 rc = fill_stats_for_pid(pid, stats);
497 if (rc < 0)
498 goto err;
499 return send_reply(rep_skb, info);
500err:
501 nlmsg_free(rep_skb);
502 return rc;
503}
504
505static int cmd_attr_tgid(struct genl_info *info)
506{
507 struct taskstats *stats;
508 struct sk_buff *rep_skb;
509 size_t size;
510 u32 tgid;
511 int rc;
512
513 size = taskstats_packet_size();
514
515 rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
516 if (rc < 0)
517 return rc;
518
519 rc = -EINVAL;
520 tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
521 stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
522 if (!stats)
523 goto err;
524
525 rc = fill_stats_for_tgid(tgid, stats);
526 if (rc < 0)
527 goto err;
528 return send_reply(rep_skb, info);
529err:
530 nlmsg_free(rep_skb);
531 return rc;
532}
533
534static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
535{
536 if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
537 return cmd_attr_cpumask(info,
538 TASKSTATS_CMD_ATTR_REGISTER_CPUMASK,
539 REGISTER);
540 else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
541 return cmd_attr_cpumask(info,
542 TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK,
543 DEREGISTER);
544 else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
545 return cmd_attr_pid(info);
546 else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
547 return cmd_attr_tgid(info);
548 else
549 return -EINVAL;
550}
551
552static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
553{
554 struct signal_struct *sig = tsk->signal;
555 struct taskstats *stats_new, *stats;
556
557 /* Pairs with smp_store_release() below. */
558 stats = smp_load_acquire(&sig->stats);
559 if (stats || thread_group_empty(tsk))
560 return stats;
561
562 /* No problem if kmem_cache_zalloc() fails */
563 stats_new = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
564
565 spin_lock_irq(&tsk->sighand->siglock);
566 stats = sig->stats;
567 if (!stats) {
568 /*
569 * Pairs with smp_store_release() above and order the
570 * kmem_cache_zalloc().
571 */
572 smp_store_release(&sig->stats, stats_new);
573 stats = stats_new;
574 stats_new = NULL;
575 }
576 spin_unlock_irq(&tsk->sighand->siglock);
577
578 if (stats_new)
579 kmem_cache_free(taskstats_cache, stats_new);
580
581 return stats;
582}
583
584/* Send pid data out on exit */
585void taskstats_exit(struct task_struct *tsk, int group_dead)
586{
587 int rc;
588 struct listener_list *listeners;
589 struct taskstats *stats;
590 struct sk_buff *rep_skb;
591 size_t size;
592 int is_thread_group;
593
594 if (!family_registered)
595 return;
596
597 /*
598 * Size includes space for nested attributes
599 */
600 size = taskstats_packet_size();
601
602 is_thread_group = !!taskstats_tgid_alloc(tsk);
603 if (is_thread_group) {
604 /* PID + STATS + TGID + STATS */
605 size = 2 * size;
606 /* fill the tsk->signal->stats structure */
607 fill_tgid_exit(tsk);
608 }
609
610 listeners = raw_cpu_ptr(&listener_array);
611 if (list_empty(&listeners->list))
612 return;
613
614 rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
615 if (rc < 0)
616 return;
617
618 stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
619 task_pid_nr_ns(tsk, &init_pid_ns));
620 if (!stats)
621 goto err;
622
623 fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
624 if (group_dead)
625 stats->ac_flag |= AGROUP;
626
627 /*
628 * Doesn't matter if tsk is the leader or the last group member leaving
629 */
630 if (!is_thread_group || !group_dead)
631 goto send;
632
633 stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
634 task_tgid_nr_ns(tsk, &init_pid_ns));
635 if (!stats)
636 goto err;
637
638 memcpy(stats, tsk->signal->stats, sizeof(*stats));
639 stats->version = TASKSTATS_VERSION;
640
641send:
642 send_cpu_listeners(rep_skb, listeners);
643 return;
644err:
645 nlmsg_free(rep_skb);
646}
647
648static const struct genl_ops taskstats_ops[] = {
649 {
650 .cmd = TASKSTATS_CMD_GET,
651 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
652 .doit = taskstats_user_cmd,
653 .policy = taskstats_cmd_get_policy,
654 .maxattr = ARRAY_SIZE(taskstats_cmd_get_policy) - 1,
655 .flags = GENL_ADMIN_PERM,
656 },
657 {
658 .cmd = CGROUPSTATS_CMD_GET,
659 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
660 .doit = cgroupstats_user_cmd,
661 .policy = cgroupstats_cmd_get_policy,
662 .maxattr = ARRAY_SIZE(cgroupstats_cmd_get_policy) - 1,
663 },
664};
665
666static struct genl_family family __ro_after_init = {
667 .name = TASKSTATS_GENL_NAME,
668 .version = TASKSTATS_GENL_VERSION,
669 .module = THIS_MODULE,
670 .ops = taskstats_ops,
671 .n_ops = ARRAY_SIZE(taskstats_ops),
672 .resv_start_op = CGROUPSTATS_CMD_GET + 1,
673 .netnsok = true,
674};
675
676/* Needed early in initialization */
677void __init taskstats_init_early(void)
678{
679 unsigned int i;
680
681 taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
682 for_each_possible_cpu(i) {
683 INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
684 init_rwsem(&(per_cpu(listener_array, i).sem));
685 }
686}
687
688static int __init taskstats_init(void)
689{
690 int rc;
691
692 rc = genl_register_family(&family);
693 if (rc)
694 return rc;
695
696 family_registered = 1;
697 pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
698 return 0;
699}
700
701/*
702 * late initcall ensures initialization of statistics collection
703 * mechanisms precedes initialization of the taskstats interface
704 */
705late_initcall(taskstats_init);