kernel/taskstats.c

   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/cpumask.h>
  24 #include <linux/percpu.h>
  25 #include <linux/slab.h>
  26 #include <linux/cgroupstats.h>
  27 #include <linux/cgroup.h>
  28 #include <linux/fs.h>
  29 #include <linux/file.h>
  30 #include <linux/pid_namespace.h>
  31 #include <net/genetlink.h>
  32 #include <linux/atomic.h>
  33
  34 /*
  35  * Maximum length of a cpumask that can be specified in
  36  * the TASKSTATS_CMD_ATTR_REGISTER/DEREGISTER_CPUMASK attribute
  37  */
  38 #define TASKSTATS_CPUMASK_MAXLEN        (100+6*NR_CPUS)
  39
  40 static DEFINE_PER_CPU(__u32, taskstats_seqnum);
  41 static int family_registered;
  42 struct kmem_cache *taskstats_cache;
  43
  44 static struct genl_family family;
  45
  46 static const struct nla_policy taskstats_cmd_get_policy[TASKSTATS_CMD_ATTR_MAX+1] = {
  47         [TASKSTATS_CMD_ATTR_PID]  = { .type = NLA_U32 },
  48         [TASKSTATS_CMD_ATTR_TGID] = { .type = NLA_U32 },
  49         [TASKSTATS_CMD_ATTR_REGISTER_CPUMASK] = { .type = NLA_STRING },
  50         [TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK] = { .type = NLA_STRING },};
  51
  52 /*
  53  * We have to use TASKSTATS_CMD_ATTR_MAX here, it is the maxattr in the family.
  54  * Make sure they are always aligned.
  55  */
  56 static const struct nla_policy cgroupstats_cmd_get_policy[TASKSTATS_CMD_ATTR_MAX+1] = {
  57         [CGROUPSTATS_CMD_ATTR_FD] = { .type = NLA_U32 },
  58 };
  59
  60 struct listener {
  61         struct list_head list;
  62         pid_t pid;
  63         char valid;
  64 };
  65
  66 struct listener_list {
  67         struct rw_semaphore sem;
  68         struct list_head list;
  69 };
  70 static DEFINE_PER_CPU(struct listener_list, listener_array);
  71
  72 enum actions {
  73         REGISTER,
  74         DEREGISTER,
  75         CPU_DONT_CARE
  76 };
  77
  78 static int prepare_reply(struct genl_info *info, u8 cmd, struct sk_buff **skbp,
  79                                 size_t size)
  80 {
  81         struct sk_buff *skb;
  82         void *reply;
  83
  84         /*
  85          * If new attributes are added, please revisit this allocation
  86          */
  87         skb = genlmsg_new(size, GFP_KERNEL);
  88         if (!skb)
  89                 return -ENOMEM;
  90
  91         if (!info) {
  92                 int seq = this_cpu_inc_return(taskstats_seqnum) - 1;
  93
  94                 reply = genlmsg_put(skb, 0, seq, &family, 0, cmd);
  95         } else
  96                 reply = genlmsg_put_reply(skb, info, &family, 0, cmd);
  97         if (reply == NULL) {
  98                 nlmsg_free(skb);
  99                 return -EINVAL;
 100         }
 101
 102         *skbp = skb;
 103         return 0;
 104 }
 105
 106 /*
 107  * Send taskstats data in @skb to listener with nl_pid @pid
 108  */
 109 static int send_reply(struct sk_buff *skb, struct genl_info *info)
 110 {
 111         struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
 112         void *reply = genlmsg_data(genlhdr);
 113
 114         genlmsg_end(skb, reply);
 115
 116         return genlmsg_reply(skb, info);
 117 }
 118
 119 /*
 120  * Send taskstats data in @skb to listeners registered for @cpu's exit data
 121  */
 122 static void send_cpu_listeners(struct sk_buff *skb,
 123                                         struct listener_list *listeners)
 124 {
 125         struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
 126         struct listener *s, *tmp;
 127         struct sk_buff *skb_next, *skb_cur = skb;
 128         void *reply = genlmsg_data(genlhdr);
 129         int rc, delcount = 0;
 130
 131         genlmsg_end(skb, reply);
 132
 133         rc = 0;
 134         down_read(&listeners->sem);
 135         list_for_each_entry(s, &listeners->list, list) {
 136                 skb_next = NULL;
 137                 if (!list_is_last(&s->list, &listeners->list)) {
 138                         skb_next = skb_clone(skb_cur, GFP_KERNEL);
 139                         if (!skb_next)
 140                                 break;
 141                 }
 142                 rc = genlmsg_unicast(&init_net, skb_cur, s->pid);
 143                 if (rc == -ECONNREFUSED) {
 144                         s->valid = 0;
 145                         delcount++;
 146                 }
 147                 skb_cur = skb_next;
 148         }
 149         up_read(&listeners->sem);
 150
 151         if (skb_cur)
 152                 nlmsg_free(skb_cur);
 153
 154         if (!delcount)
 155                 return;
 156
 157         /* Delete invalidated entries */
 158         down_write(&listeners->sem);
 159         list_for_each_entry_safe(s, tmp, &listeners->list, list) {
 160                 if (!s->valid) {
 161                         list_del(&s->list);
 162                         kfree(s);
 163                 }
 164         }
 165         up_write(&listeners->sem);
 166 }
 167
 168 static void fill_stats(struct user_namespace *user_ns,
 169                        struct pid_namespace *pid_ns,
 170                        struct task_struct *tsk, struct taskstats *stats)
 171 {
 172         memset(stats, 0, sizeof(*stats));
 173         /*
 174          * Each accounting subsystem adds calls to its functions to
 175          * fill in relevant parts of struct taskstsats as follows
 176          *
 177          *      per-task-foo(stats, tsk);
 178          */
 179
 180         delayacct_add_tsk(stats, tsk);
 181
 182         /* fill in basic acct fields */
 183         stats->version = TASKSTATS_VERSION;
 184         stats->nvcsw = tsk->nvcsw;
 185         stats->nivcsw = tsk->nivcsw;
 186         bacct_add_tsk(user_ns, pid_ns, stats, tsk);
 187
 188         /* fill in extended acct fields */
 189         xacct_add_tsk(stats, tsk);
 190 }
 191
 192 static int fill_stats_for_pid(pid_t pid, struct taskstats *stats)
 193 {
 194         struct task_struct *tsk;
 195
 196         rcu_read_lock();
 197         tsk = find_task_by_vpid(pid);
 198         if (tsk)
 199                 get_task_struct(tsk);
 200         rcu_read_unlock();
 201         if (!tsk)
 202                 return -ESRCH;
 203         fill_stats(current_user_ns(), task_active_pid_ns(current), tsk, stats);
 204         put_task_struct(tsk);
 205         return 0;
 206 }
 207
 208 static int fill_stats_for_tgid(pid_t tgid, struct taskstats *stats)
 209 {
 210         struct task_struct *tsk, *first;
 211         unsigned long flags;
 212         int rc = -ESRCH;
 213
 214         /*
 215          * Add additional stats from live tasks except zombie thread group
 216          * leaders who are already counted with the dead tasks
 217          */
 218         rcu_read_lock();
 219         first = find_task_by_vpid(tgid);
 220
 221         if (!first || !lock_task_sighand(first, &flags))
 222                 goto out;
 223
 224         if (first->signal->stats)
 225                 memcpy(stats, first->signal->stats, sizeof(*stats));
 226         else
 227                 memset(stats, 0, sizeof(*stats));
 228
 229         tsk = first;
 230         do {
 231                 if (tsk->exit_state)
 232                         continue;
 233                 /*
 234                  * Accounting subsystem can call its functions here to
 235                  * fill in relevant parts of struct taskstsats as follows
 236                  *
 237                  *      per-task-foo(stats, tsk);
 238                  */
 239                 delayacct_add_tsk(stats, tsk);
 240
 241                 stats->nvcsw += tsk->nvcsw;
 242                 stats->nivcsw += tsk->nivcsw;
 243         } while_each_thread(first, tsk);
 244
 245         unlock_task_sighand(first, &flags);
 246         rc = 0;
 247 out:
 248         rcu_read_unlock();
 249
 250         stats->version = TASKSTATS_VERSION;
 251         /*
 252          * Accounting subsystems can also add calls here to modify
 253          * fields of taskstats.
 254          */
 255         return rc;
 256 }
 257
 258 static void fill_tgid_exit(struct task_struct *tsk)
 259 {
 260         unsigned long flags;
 261
 262         spin_lock_irqsave(&tsk->sighand->siglock, flags);
 263         if (!tsk->signal->stats)
 264                 goto ret;
 265
 266         /*
 267          * Each accounting subsystem calls its functions here to
 268          * accumalate its per-task stats for tsk, into the per-tgid structure
 269          *
 270          *      per-task-foo(tsk->signal->stats, tsk);
 271          */
 272         delayacct_add_tsk(tsk->signal->stats, tsk);
 273 ret:
 274         spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
 275         return;
 276 }
 277
 278 static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
 279 {
 280         struct listener_list *listeners;
 281         struct listener *s, *tmp, *s2;
 282         unsigned int cpu;
 283         int ret = 0;
 284
 285         if (!cpumask_subset(mask, cpu_possible_mask))
 286                 return -EINVAL;
 287
 288         if (current_user_ns() != &init_user_ns)
 289                 return -EINVAL;
 290
 291         if (task_active_pid_ns(current) != &init_pid_ns)
 292                 return -EINVAL;
 293
 294         if (isadd == REGISTER) {
 295                 for_each_cpu(cpu, mask) {
 296                         s = kmalloc_node(sizeof(struct listener),
 297                                         GFP_KERNEL, cpu_to_node(cpu));
 298                         if (!s) {
 299                                 ret = -ENOMEM;
 300                                 goto cleanup;
 301                         }
 302                         s->pid = pid;
 303                         s->valid = 1;
 304
 305                         listeners = &per_cpu(listener_array, cpu);
 306                         down_write(&listeners->sem);
 307                         list_for_each_entry(s2, &listeners->list, list) {
 308                                 if (s2->pid == pid && s2->valid)
 309                                         goto exists;
 310                         }
 311                         list_add(&s->list, &listeners->list);
 312                         s = NULL;
 313 exists:
 314                         up_write(&listeners->sem);
 315                         kfree(s); /* nop if NULL */
 316                 }
 317                 return 0;
 318         }
 319
 320         /* Deregister or cleanup */
 321 cleanup:
 322         for_each_cpu(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 ret;
 335 }
 336
 337 static int parse(struct nlattr *na, struct cpumask *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
 359 static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
 360 {
 361         struct nlattr *na, *ret;
 362         int aggr;
 363
 364         aggr = (type == TASKSTATS_TYPE_PID)
 365                         ? TASKSTATS_TYPE_AGGR_PID
 366                         : TASKSTATS_TYPE_AGGR_TGID;
 367
 368         na = nla_nest_start(skb, aggr);
 369         if (!na)
 370                 goto err;
 371
 372         if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
 373                 nla_nest_cancel(skb, na);
 374                 goto err;
 375         }
 376         ret = nla_reserve_64bit(skb, TASKSTATS_TYPE_STATS,
 377                                 sizeof(struct taskstats), TASKSTATS_TYPE_NULL);
 378         if (!ret) {
 379                 nla_nest_cancel(skb, na);
 380                 goto err;
 381         }
 382         nla_nest_end(skb, na);
 383
 384         return nla_data(ret);
 385 err:
 386         return NULL;
 387 }
 388
 389 static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
 390 {
 391         int rc = 0;
 392         struct sk_buff *rep_skb;
 393         struct cgroupstats *stats;
 394         struct nlattr *na;
 395         size_t size;
 396         u32 fd;
 397         struct fd f;
 398
 399         na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
 400         if (!na)
 401                 return -EINVAL;
 402
 403         fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
 404         f = fdget(fd);
 405         if (!f.file)
 406                 return 0;
 407
 408         size = nla_total_size(sizeof(struct cgroupstats));
 409
 410         rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
 411                                 size);
 412         if (rc < 0)
 413                 goto err;
 414
 415         na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
 416                                 sizeof(struct cgroupstats));
 417         if (na == NULL) {
 418                 nlmsg_free(rep_skb);
 419                 rc = -EMSGSIZE;
 420                 goto err;
 421         }
 422
 423         stats = nla_data(na);
 424         memset(stats, 0, sizeof(*stats));
 425
 426         rc = cgroupstats_build(stats, f.file->f_path.dentry);
 427         if (rc < 0) {
 428                 nlmsg_free(rep_skb);
 429                 goto err;
 430         }
 431
 432         rc = send_reply(rep_skb, info);
 433
 434 err:
 435         fdput(f);
 436         return rc;
 437 }
 438
 439 static int cmd_attr_register_cpumask(struct genl_info *info)
 440 {
 441         cpumask_var_t mask;
 442         int rc;
 443
 444         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
 445                 return -ENOMEM;
 446         rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
 447         if (rc < 0)
 448                 goto out;
 449         rc = add_del_listener(info->snd_portid, mask, REGISTER);
 450 out:
 451         free_cpumask_var(mask);
 452         return rc;
 453 }
 454
 455 static int cmd_attr_deregister_cpumask(struct genl_info *info)
 456 {
 457         cpumask_var_t mask;
 458         int rc;
 459
 460         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
 461                 return -ENOMEM;
 462         rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
 463         if (rc < 0)
 464                 goto out;
 465         rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
 466 out:
 467         free_cpumask_var(mask);
 468         return rc;
 469 }
 470
 471 static size_t taskstats_packet_size(void)
 472 {
 473         size_t size;
 474
 475         size = nla_total_size(sizeof(u32)) +
 476                 nla_total_size_64bit(sizeof(struct taskstats)) +
 477                 nla_total_size(0);
 478
 479         return size;
 480 }
 481
 482 static int cmd_attr_pid(struct genl_info *info)
 483 {
 484         struct taskstats *stats;
 485         struct sk_buff *rep_skb;
 486         size_t size;
 487         u32 pid;
 488         int rc;
 489
 490         size = taskstats_packet_size();
 491
 492         rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
 493         if (rc < 0)
 494                 return rc;
 495
 496         rc = -EINVAL;
 497         pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
 498         stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
 499         if (!stats)
 500                 goto err;
 501
 502         rc = fill_stats_for_pid(pid, stats);
 503         if (rc < 0)
 504                 goto err;
 505         return send_reply(rep_skb, info);
 506 err:
 507         nlmsg_free(rep_skb);
 508         return rc;
 509 }
 510
 511 static int cmd_attr_tgid(struct genl_info *info)
 512 {
 513         struct taskstats *stats;
 514         struct sk_buff *rep_skb;
 515         size_t size;
 516         u32 tgid;
 517         int rc;
 518
 519         size = taskstats_packet_size();
 520
 521         rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
 522         if (rc < 0)
 523                 return rc;
 524
 525         rc = -EINVAL;
 526         tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
 527         stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
 528         if (!stats)
 529                 goto err;
 530
 531         rc = fill_stats_for_tgid(tgid, stats);
 532         if (rc < 0)
 533                 goto err;
 534         return send_reply(rep_skb, info);
 535 err:
 536         nlmsg_free(rep_skb);
 537         return rc;
 538 }
 539
 540 static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
 541 {
 542         if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
 543                 return cmd_attr_register_cpumask(info);
 544         else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
 545                 return cmd_attr_deregister_cpumask(info);
 546         else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
 547                 return cmd_attr_pid(info);
 548         else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
 549                 return cmd_attr_tgid(info);
 550         else
 551                 return -EINVAL;
 552 }
 553
 554 static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
 555 {
 556         struct signal_struct *sig = tsk->signal;
 557         struct taskstats *stats;
 558
 559         if (sig->stats || thread_group_empty(tsk))
 560                 goto ret;
 561
 562         /* No problem if kmem_cache_zalloc() fails */
 563         stats = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
 564
 565         spin_lock_irq(&tsk->sighand->siglock);
 566         if (!sig->stats) {
 567                 sig->stats = stats;
 568                 stats = NULL;
 569         }
 570         spin_unlock_irq(&tsk->sighand->siglock);
 571
 572         if (stats)
 573                 kmem_cache_free(taskstats_cache, stats);
 574 ret:
 575         return sig->stats;
 576 }
 577
 578 /* Send pid data out on exit */
 579 void taskstats_exit(struct task_struct *tsk, int group_dead)
 580 {
 581         int rc;
 582         struct listener_list *listeners;
 583         struct taskstats *stats;
 584         struct sk_buff *rep_skb;
 585         size_t size;
 586         int is_thread_group;
 587
 588         if (!family_registered)
 589                 return;
 590
 591         /*
 592          * Size includes space for nested attributes
 593          */
 594         size = taskstats_packet_size();
 595
 596         is_thread_group = !!taskstats_tgid_alloc(tsk);
 597         if (is_thread_group) {
 598                 /* PID + STATS + TGID + STATS */
 599                 size = 2 * size;
 600                 /* fill the tsk->signal->stats structure */
 601                 fill_tgid_exit(tsk);
 602         }
 603
 604         listeners = raw_cpu_ptr(&listener_array);
 605         if (list_empty(&listeners->list))
 606                 return;
 607
 608         rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
 609         if (rc < 0)
 610                 return;
 611
 612         stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
 613                          task_pid_nr_ns(tsk, &init_pid_ns));
 614         if (!stats)
 615                 goto err;
 616
 617         fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
 618
 619         /*
 620          * Doesn't matter if tsk is the leader or the last group member leaving
 621          */
 622         if (!is_thread_group || !group_dead)
 623                 goto send;
 624
 625         stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
 626                          task_tgid_nr_ns(tsk, &init_pid_ns));
 627         if (!stats)
 628                 goto err;
 629
 630         memcpy(stats, tsk->signal->stats, sizeof(*stats));
 631
 632 send:
 633         send_cpu_listeners(rep_skb, listeners);
 634         return;
 635 err:
 636         nlmsg_free(rep_skb);
 637 }
 638
 639 static const struct genl_ops taskstats_ops[] = {
 640         {
 641                 .cmd            = TASKSTATS_CMD_GET,
 642                 .doit           = taskstats_user_cmd,
 643                 .policy         = taskstats_cmd_get_policy,
 644                 .flags          = GENL_ADMIN_PERM,
 645         },
 646         {
 647                 .cmd            = CGROUPSTATS_CMD_GET,
 648                 .doit           = cgroupstats_user_cmd,
 649                 .policy         = cgroupstats_cmd_get_policy,
 650         },
 651 };
 652
 653 static struct genl_family family __ro_after_init = {
 654         .name           = TASKSTATS_GENL_NAME,
 655         .version        = TASKSTATS_GENL_VERSION,
 656         .maxattr        = TASKSTATS_CMD_ATTR_MAX,
 657         .module         = THIS_MODULE,
 658         .ops            = taskstats_ops,
 659         .n_ops          = ARRAY_SIZE(taskstats_ops),
 660 };
 661
 662 /* Needed early in initialization */
 663 void __init taskstats_init_early(void)
 664 {
 665         unsigned int i;
 666
 667         taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
 668         for_each_possible_cpu(i) {
 669                 INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
 670                 init_rwsem(&(per_cpu(listener_array, i).sem));
 671         }
 672 }
 673
 674 static int __init taskstats_init(void)
 675 {
 676         int rc;
 677
 678         rc = genl_register_family(&family);
 679         if (rc)
 680                 return rc;
 681
 682         family_registered = 1;
 683         pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
 684         return 0;
 685 }
 686
 687 /*
 688  * late initcall ensures initialization of statistics collection
 689  * mechanisms precedes initialization of the taskstats interface
 690  */
 691 late_initcall(taskstats_init);