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(payload: 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, portid: 0, seq, family: &family, flags: 0, cmd);
83 } else
84 reply = genlmsg_put_reply(skb, info, family: &family, flags: 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(nlh: nlmsg_hdr(skb));
100 void *reply = genlmsg_data(gnlh: genlhdr);
101
102 genlmsg_end(skb, hdr: 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(nlh: nlmsg_hdr(skb));
114 struct listener *s, *tmp;
115 struct sk_buff *skb_next, *skb_cur = skb;
116 void *reply = genlmsg_data(gnlh: genlhdr);
117 int delcount = 0;
118
119 genlmsg_end(skb, hdr: reply);
120
121 down_read(sem: &listeners->sem);
122 list_for_each_entry(s, &listeners->list, list) {
123 int rc;
124
125 skb_next = NULL;
126 if (!list_is_last(list: &s->list, head: &listeners->list)) {
127 skb_next = skb_clone(skb: skb_cur, GFP_KERNEL);
128 if (!skb_next)
129 break;
130 }
131 rc = genlmsg_unicast(net: &init_net, skb: skb_cur, portid: s->pid);
132 if (rc == -ECONNREFUSED) {
133 s->valid = 0;
134 delcount++;
135 }
136 skb_cur = skb_next;
137 }
138 up_read(sem: &listeners->sem);
139
140 if (skb_cur)
141 nlmsg_free(skb: skb_cur);
142
143 if (!delcount)
144 return;
145
146 /* Delete invalidated entries */
147 down_write(sem: &listeners->sem);
148 list_for_each_entry_safe(s, tmp, &listeners->list, list) {
149 if (!s->valid) {
150 list_del(entry: &s->list);
151 kfree(objp: s);
152 }
153 }
154 up_write(sem: &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(task: tsk);
161
162 if (exe_file) {
163 /* Following cp_new_stat64() in stat.c . */
164 stats->ac_exe_dev =
165 huge_encode_dev(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(s: stats, c: 0, n: 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, p: 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(nr: pid);
206 if (!tsk)
207 return -ESRCH;
208 fill_stats(user_ns: current_user_ns(), pid_ns: task_active_pid_ns(current), tsk, stats);
209 put_task_struct(t: 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(nr: tgid);
227
228 if (!first || !lock_task_sighand(task: first, flags: &flags))
229 goto out;
230
231 if (first->signal->stats)
232 memcpy(to: stats, from: first->signal->stats, len: sizeof(*stats));
233 else
234 memset(s: stats, c: 0, n: sizeof(*stats));
235
236 start_time = ktime_get_ns();
237 for_each_thread(first, tsk) {
238 if (tsk->exit_state)
239 continue;
240 /*
241 * Accounting subsystem can call its functions here to
242 * fill in relevant parts of struct taskstsats as follows
243 *
244 * per-task-foo(stats, tsk);
245 */
246 delayacct_add_tsk(stats, tsk);
247
248 /* calculate task elapsed time in nsec */
249 delta = start_time - tsk->start_time;
250 /* Convert to micro seconds */
251 do_div(delta, NSEC_PER_USEC);
252 stats->ac_etime += delta;
253
254 task_cputime(t: tsk, utime: &utime, stime: &stime);
255 stats->ac_utime += div_u64(dividend: utime, NSEC_PER_USEC);
256 stats->ac_stime += div_u64(dividend: stime, NSEC_PER_USEC);
257
258 stats->nvcsw += tsk->nvcsw;
259 stats->nivcsw += tsk->nivcsw;
260 }
261
262 unlock_task_sighand(task: first, flags: &flags);
263 rc = 0;
264out:
265 rcu_read_unlock();
266
267 stats->version = TASKSTATS_VERSION;
268 /*
269 * Accounting subsystems can also add calls here to modify
270 * fields of taskstats.
271 */
272 return rc;
273}
274
275static void fill_tgid_exit(struct task_struct *tsk)
276{
277 unsigned long flags;
278
279 spin_lock_irqsave(&tsk->sighand->siglock, flags);
280 if (!tsk->signal->stats)
281 goto ret;
282
283 /*
284 * Each accounting subsystem calls its functions here to
285 * accumalate its per-task stats for tsk, into the per-tgid structure
286 *
287 * per-task-foo(tsk->signal->stats, tsk);
288 */
289 delayacct_add_tsk(tsk->signal->stats, tsk);
290ret:
291 spin_unlock_irqrestore(lock: &tsk->sighand->siglock, flags);
292 return;
293}
294
295static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
296{
297 struct listener_list *listeners;
298 struct listener *s, *tmp, *s2;
299 unsigned int cpu;
300 int ret = 0;
301
302 if (!cpumask_subset(src1p: mask, cpu_possible_mask))
303 return -EINVAL;
304
305 if (current_user_ns() != &init_user_ns)
306 return -EINVAL;
307
308 if (task_active_pid_ns(current) != &init_pid_ns)
309 return -EINVAL;
310
311 if (isadd == REGISTER) {
312 for_each_cpu(cpu, mask) {
313 s = kmalloc_node(sizeof(struct listener),
314 GFP_KERNEL, cpu_to_node(cpu));
315 if (!s) {
316 ret = -ENOMEM;
317 goto cleanup;
318 }
319 s->pid = pid;
320 s->valid = 1;
321
322 listeners = &per_cpu(listener_array, cpu);
323 down_write(sem: &listeners->sem);
324 list_for_each_entry(s2, &listeners->list, list) {
325 if (s2->pid == pid && s2->valid)
326 goto exists;
327 }
328 list_add(new: &s->list, head: &listeners->list);
329 s = NULL;
330exists:
331 up_write(sem: &listeners->sem);
332 kfree(objp: s); /* nop if NULL */
333 }
334 return 0;
335 }
336
337 /* Deregister or cleanup */
338cleanup:
339 for_each_cpu(cpu, mask) {
340 listeners = &per_cpu(listener_array, cpu);
341 down_write(sem: &listeners->sem);
342 list_for_each_entry_safe(s, tmp, &listeners->list, list) {
343 if (s->pid == pid) {
344 list_del(entry: &s->list);
345 kfree(objp: s);
346 break;
347 }
348 }
349 up_write(sem: &listeners->sem);
350 }
351 return ret;
352}
353
354static int parse(struct nlattr *na, struct cpumask *mask)
355{
356 char *data;
357 int len;
358 int ret;
359
360 if (na == NULL)
361 return 1;
362 len = nla_len(nla: na);
363 if (len > TASKSTATS_CPUMASK_MAXLEN)
364 return -E2BIG;
365 if (len < 1)
366 return -EINVAL;
367 data = kmalloc(len, GFP_KERNEL);
368 if (!data)
369 return -ENOMEM;
370 nla_strscpy(dst: data, nla: na, dstsize: len);
371 ret = cpulist_parse(buf: data, dstp: mask);
372 kfree(objp: data);
373 return ret;
374}
375
376static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
377{
378 struct nlattr *na, *ret;
379 int aggr;
380
381 aggr = (type == TASKSTATS_TYPE_PID)
382 ? TASKSTATS_TYPE_AGGR_PID
383 : TASKSTATS_TYPE_AGGR_TGID;
384
385 na = nla_nest_start_noflag(skb, attrtype: aggr);
386 if (!na)
387 goto err;
388
389 if (nla_put(skb, attrtype: type, attrlen: sizeof(pid), data: &pid) < 0) {
390 nla_nest_cancel(skb, start: na);
391 goto err;
392 }
393 ret = nla_reserve_64bit(skb, attrtype: TASKSTATS_TYPE_STATS,
394 attrlen: sizeof(struct taskstats), padattr: TASKSTATS_TYPE_NULL);
395 if (!ret) {
396 nla_nest_cancel(skb, start: na);
397 goto err;
398 }
399 nla_nest_end(skb, start: na);
400
401 return nla_data(nla: ret);
402err:
403 return NULL;
404}
405
406static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
407{
408 int rc = 0;
409 struct sk_buff *rep_skb;
410 struct cgroupstats *stats;
411 struct nlattr *na;
412 size_t size;
413 u32 fd;
414
415 na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
416 if (!na)
417 return -EINVAL;
418
419 fd = nla_get_u32(nla: info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
420 CLASS(fd, f)(fd);
421 if (fd_empty(f))
422 return 0;
423
424 size = nla_total_size(payload: sizeof(struct cgroupstats));
425
426 rc = prepare_reply(info, cmd: CGROUPSTATS_CMD_NEW, skbp: &rep_skb,
427 size);
428 if (rc < 0)
429 return rc;
430
431 na = nla_reserve(skb: rep_skb, attrtype: CGROUPSTATS_TYPE_CGROUP_STATS,
432 attrlen: sizeof(struct cgroupstats));
433 if (na == NULL) {
434 nlmsg_free(skb: rep_skb);
435 return -EMSGSIZE;
436 }
437
438 stats = nla_data(nla: na);
439 memset(s: stats, c: 0, n: sizeof(*stats));
440
441 rc = cgroupstats_build(stats, fd_file(f)->f_path.dentry);
442 if (rc < 0) {
443 nlmsg_free(skb: rep_skb);
444 return rc;
445 }
446
447 return send_reply(skb: rep_skb, info);
448}
449
450static int cmd_attr_register_cpumask(struct genl_info *info)
451{
452 cpumask_var_t mask;
453 int rc;
454
455 if (!alloc_cpumask_var(mask: &mask, GFP_KERNEL))
456 return -ENOMEM;
457 rc = parse(na: info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
458 if (rc < 0)
459 goto out;
460 rc = add_del_listener(pid: info->snd_portid, mask, isadd: REGISTER);
461out:
462 free_cpumask_var(mask);
463 return rc;
464}
465
466static int cmd_attr_deregister_cpumask(struct genl_info *info)
467{
468 cpumask_var_t mask;
469 int rc;
470
471 if (!alloc_cpumask_var(mask: &mask, GFP_KERNEL))
472 return -ENOMEM;
473 rc = parse(na: info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
474 if (rc < 0)
475 goto out;
476 rc = add_del_listener(pid: info->snd_portid, mask, isadd: DEREGISTER);
477out:
478 free_cpumask_var(mask);
479 return rc;
480}
481
482static size_t taskstats_packet_size(void)
483{
484 size_t size;
485
486 size = nla_total_size(payload: sizeof(u32)) +
487 nla_total_size_64bit(payload: sizeof(struct taskstats)) +
488 nla_total_size(payload: 0);
489
490 return size;
491}
492
493static int cmd_attr_pid(struct genl_info *info)
494{
495 struct taskstats *stats;
496 struct sk_buff *rep_skb;
497 size_t size;
498 u32 pid;
499 int rc;
500
501 size = taskstats_packet_size();
502
503 rc = prepare_reply(info, cmd: TASKSTATS_CMD_NEW, skbp: &rep_skb, size);
504 if (rc < 0)
505 return rc;
506
507 rc = -EINVAL;
508 pid = nla_get_u32(nla: info->attrs[TASKSTATS_CMD_ATTR_PID]);
509 stats = mk_reply(skb: rep_skb, type: TASKSTATS_TYPE_PID, pid);
510 if (!stats)
511 goto err;
512
513 rc = fill_stats_for_pid(pid, stats);
514 if (rc < 0)
515 goto err;
516 return send_reply(skb: rep_skb, info);
517err:
518 nlmsg_free(skb: rep_skb);
519 return rc;
520}
521
522static int cmd_attr_tgid(struct genl_info *info)
523{
524 struct taskstats *stats;
525 struct sk_buff *rep_skb;
526 size_t size;
527 u32 tgid;
528 int rc;
529
530 size = taskstats_packet_size();
531
532 rc = prepare_reply(info, cmd: TASKSTATS_CMD_NEW, skbp: &rep_skb, size);
533 if (rc < 0)
534 return rc;
535
536 rc = -EINVAL;
537 tgid = nla_get_u32(nla: info->attrs[TASKSTATS_CMD_ATTR_TGID]);
538 stats = mk_reply(skb: rep_skb, type: TASKSTATS_TYPE_TGID, pid: tgid);
539 if (!stats)
540 goto err;
541
542 rc = fill_stats_for_tgid(tgid, stats);
543 if (rc < 0)
544 goto err;
545 return send_reply(skb: rep_skb, info);
546err:
547 nlmsg_free(skb: rep_skb);
548 return rc;
549}
550
551static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
552{
553 if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
554 return cmd_attr_register_cpumask(info);
555 else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
556 return cmd_attr_deregister_cpumask(info);
557 else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
558 return cmd_attr_pid(info);
559 else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
560 return cmd_attr_tgid(info);
561 else
562 return -EINVAL;
563}
564
565static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
566{
567 struct signal_struct *sig = tsk->signal;
568 struct taskstats *stats_new, *stats;
569
570 /* Pairs with smp_store_release() below. */
571 stats = smp_load_acquire(&sig->stats);
572 if (stats || thread_group_empty(p: tsk))
573 return stats;
574
575 /* No problem if kmem_cache_zalloc() fails */
576 stats_new = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
577
578 spin_lock_irq(lock: &tsk->sighand->siglock);
579 stats = sig->stats;
580 if (!stats) {
581 /*
582 * Pairs with smp_store_release() above and order the
583 * kmem_cache_zalloc().
584 */
585 smp_store_release(&sig->stats, stats_new);
586 stats = stats_new;
587 stats_new = NULL;
588 }
589 spin_unlock_irq(lock: &tsk->sighand->siglock);
590
591 if (stats_new)
592 kmem_cache_free(s: taskstats_cache, objp: stats_new);
593
594 return stats;
595}
596
597/* Send pid data out on exit */
598void taskstats_exit(struct task_struct *tsk, int group_dead)
599{
600 int rc;
601 struct listener_list *listeners;
602 struct taskstats *stats;
603 struct sk_buff *rep_skb;
604 size_t size;
605 int is_thread_group;
606
607 if (!family_registered)
608 return;
609
610 /*
611 * Size includes space for nested attributes
612 */
613 size = taskstats_packet_size();
614
615 is_thread_group = !!taskstats_tgid_alloc(tsk);
616 if (is_thread_group) {
617 /* PID + STATS + TGID + STATS */
618 size = 2 * size;
619 /* fill the tsk->signal->stats structure */
620 fill_tgid_exit(tsk);
621 }
622
623 listeners = raw_cpu_ptr(&listener_array);
624 if (list_empty(head: &listeners->list))
625 return;
626
627 rc = prepare_reply(NULL, cmd: TASKSTATS_CMD_NEW, skbp: &rep_skb, size);
628 if (rc < 0)
629 return;
630
631 stats = mk_reply(skb: rep_skb, type: TASKSTATS_TYPE_PID,
632 pid: task_pid_nr_ns(tsk, ns: &init_pid_ns));
633 if (!stats)
634 goto err;
635
636 fill_stats(user_ns: &init_user_ns, pid_ns: &init_pid_ns, tsk, stats);
637 if (group_dead)
638 stats->ac_flag |= AGROUP;
639
640 /*
641 * Doesn't matter if tsk is the leader or the last group member leaving
642 */
643 if (!is_thread_group || !group_dead)
644 goto send;
645
646 stats = mk_reply(skb: rep_skb, type: TASKSTATS_TYPE_TGID,
647 pid: task_tgid_nr_ns(tsk, ns: &init_pid_ns));
648 if (!stats)
649 goto err;
650
651 memcpy(to: stats, from: tsk->signal->stats, len: sizeof(*stats));
652
653send:
654 send_cpu_listeners(skb: rep_skb, listeners);
655 return;
656err:
657 nlmsg_free(skb: rep_skb);
658}
659
660static const struct genl_ops taskstats_ops[] = {
661 {
662 .cmd = TASKSTATS_CMD_GET,
663 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
664 .doit = taskstats_user_cmd,
665 .policy = taskstats_cmd_get_policy,
666 .maxattr = ARRAY_SIZE(taskstats_cmd_get_policy) - 1,
667 .flags = GENL_ADMIN_PERM,
668 },
669 {
670 .cmd = CGROUPSTATS_CMD_GET,
671 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
672 .doit = cgroupstats_user_cmd,
673 .policy = cgroupstats_cmd_get_policy,
674 .maxattr = ARRAY_SIZE(cgroupstats_cmd_get_policy) - 1,
675 },
676};
677
678static struct genl_family family __ro_after_init = {
679 .name = TASKSTATS_GENL_NAME,
680 .version = TASKSTATS_GENL_VERSION,
681 .module = THIS_MODULE,
682 .ops = taskstats_ops,
683 .n_ops = ARRAY_SIZE(taskstats_ops),
684 .resv_start_op = CGROUPSTATS_CMD_GET + 1,
685 .netnsok = true,
686};
687
688/* Needed early in initialization */
689void __init taskstats_init_early(void)
690{
691 unsigned int i;
692
693 taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
694 for_each_possible_cpu(i) {
695 INIT_LIST_HEAD(list: &(per_cpu(listener_array, i).list));
696 init_rwsem(&(per_cpu(listener_array, i).sem));
697 }
698}
699
700static int __init taskstats_init(void)
701{
702 int rc;
703
704 rc = genl_register_family(family: &family);
705 if (rc)
706 return rc;
707
708 family_registered = 1;
709 pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
710 return 0;
711}
712
713/*
714 * late initcall ensures initialization of statistics collection
715 * mechanisms precedes initialization of the taskstats interface
716 */
717late_initcall(taskstats_init);
718