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         .id             = GENL_ID_GENERATE,
  46         .name           = TASKSTATS_GENL_NAME,
  47         .version        = TASKSTATS_GENL_VERSION,
  48         .maxattr        = TASKSTATS_CMD_ATTR_MAX,
  49 };
  50
  51 static const struct nla_policy taskstats_cmd_get_policy[TASKSTATS_CMD_ATTR_MAX+1] = {
  52         [TASKSTATS_CMD_ATTR_PID]  = { .type = NLA_U32 },
  53         [TASKSTATS_CMD_ATTR_TGID] = { .type = NLA_U32 },
  54         [TASKSTATS_CMD_ATTR_REGISTER_CPUMASK] = { .type = NLA_STRING },
  55         [TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK] = { .type = NLA_STRING },};
  56
  57 /*
  58  * We have to use TASKSTATS_CMD_ATTR_MAX here, it is the maxattr in the family.
  59  * Make sure they are always aligned.
  60  */
  61 static const struct nla_policy cgroupstats_cmd_get_policy[TASKSTATS_CMD_ATTR_MAX+1] = {
  62         [CGROUPSTATS_CMD_ATTR_FD] = { .type = NLA_U32 },
  63 };
  64
  65 struct listener {
  66         struct list_head list;
  67         pid_t pid;
  68         char valid;
  69 };
  70
  71 struct listener_list {
  72         struct rw_semaphore sem;
  73         struct list_head list;
  74 };
  75 static DEFINE_PER_CPU(struct listener_list, listener_array);
  76
  77 enum actions {
  78         REGISTER,
  79         DEREGISTER,
  80         CPU_DONT_CARE
  81 };
  82
  83 static int prepare_reply(struct genl_info *info, u8 cmd, struct sk_buff **skbp,
  84                                 size_t size)
  85 {
  86         struct sk_buff *skb;
  87         void *reply;
  88
  89         /*
  90          * If new attributes are added, please revisit this allocation
  91          */
  92         skb = genlmsg_new(size, GFP_KERNEL);
  93         if (!skb)
  94                 return -ENOMEM;
  95
  96         if (!info) {
  97                 int seq = this_cpu_inc_return(taskstats_seqnum) - 1;
  98
  99                 reply = genlmsg_put(skb, 0, seq, &family, 0, cmd);
 100         } else
 101                 reply = genlmsg_put_reply(skb, info, &family, 0, cmd);
 102         if (reply == NULL) {
 103                 nlmsg_free(skb);
 104                 return -EINVAL;
 105         }
 106
 107         *skbp = skb;
 108         return 0;
 109 }
 110
 111 /*
 112  * Send taskstats data in @skb to listener with nl_pid @pid
 113  */
 114 static int send_reply(struct sk_buff *skb, struct genl_info *info)
 115 {
 116         struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
 117         void *reply = genlmsg_data(genlhdr);
 118
 119         genlmsg_end(skb, reply);
 120
 121         return genlmsg_reply(skb, info);
 122 }
 123
 124 /*
 125  * Send taskstats data in @skb to listeners registered for @cpu's exit data
 126  */
 127 static void send_cpu_listeners(struct sk_buff *skb,
 128                                         struct listener_list *listeners)
 129 {
 130         struct genlmsghdr *genlhdr = nlmsg_data(nlmsg_hdr(skb));
 131         struct listener *s, *tmp;
 132         struct sk_buff *skb_next, *skb_cur = skb;
 133         void *reply = genlmsg_data(genlhdr);
 134         int rc, delcount = 0;
 135
 136         genlmsg_end(skb, reply);
 137
 138         rc = 0;
 139         down_read(&listeners->sem);
 140         list_for_each_entry(s, &listeners->list, list) {
 141                 skb_next = NULL;
 142                 if (!list_is_last(&s->list, &listeners->list)) {
 143                         skb_next = skb_clone(skb_cur, GFP_KERNEL);
 144                         if (!skb_next)
 145                                 break;
 146                 }
 147                 rc = genlmsg_unicast(&init_net, skb_cur, s->pid);
 148                 if (rc == -ECONNREFUSED) {
 149                         s->valid = 0;
 150                         delcount++;
 151                 }
 152                 skb_cur = skb_next;
 153         }
 154         up_read(&listeners->sem);
 155
 156         if (skb_cur)
 157                 nlmsg_free(skb_cur);
 158
 159         if (!delcount)
 160                 return;
 161
 162         /* Delete invalidated entries */
 163         down_write(&listeners->sem);
 164         list_for_each_entry_safe(s, tmp, &listeners->list, list) {
 165                 if (!s->valid) {
 166                         list_del(&s->list);
 167                         kfree(s);
 168                 }
 169         }
 170         up_write(&listeners->sem);
 171 }
 172
 173 static void fill_stats(struct user_namespace *user_ns,
 174                        struct pid_namespace *pid_ns,
 175                        struct task_struct *tsk, struct taskstats *stats)
 176 {
 177         memset(stats, 0, sizeof(*stats));
 178         /*
 179          * Each accounting subsystem adds calls to its functions to
 180          * fill in relevant parts of struct taskstsats as follows
 181          *
 182          *      per-task-foo(stats, tsk);
 183          */
 184
 185         delayacct_add_tsk(stats, tsk);
 186
 187         /* fill in basic acct fields */
 188         stats->version = TASKSTATS_VERSION;
 189         stats->nvcsw = tsk->nvcsw;
 190         stats->nivcsw = tsk->nivcsw;
 191         bacct_add_tsk(user_ns, pid_ns, stats, tsk);
 192
 193         /* fill in extended acct fields */
 194         xacct_add_tsk(stats, tsk);
 195 }
 196
 197 static int fill_stats_for_pid(pid_t pid, struct taskstats *stats)
 198 {
 199         struct task_struct *tsk;
 200
 201         rcu_read_lock();
 202         tsk = find_task_by_vpid(pid);
 203         if (tsk)
 204                 get_task_struct(tsk);
 205         rcu_read_unlock();
 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
 213 static 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
 219         /*
 220          * Add additional stats from live tasks except zombie thread group
 221          * leaders who are already counted with the dead tasks
 222          */
 223         rcu_read_lock();
 224         first = find_task_by_vpid(tgid);
 225
 226         if (!first || !lock_task_sighand(first, &flags))
 227                 goto out;
 228
 229         if (first->signal->stats)
 230                 memcpy(stats, first->signal->stats, sizeof(*stats));
 231         else
 232                 memset(stats, 0, sizeof(*stats));
 233
 234         tsk = first;
 235         do {
 236                 if (tsk->exit_state)
 237                         continue;
 238                 /*
 239                  * Accounting subsystem can call its functions here to
 240                  * fill in relevant parts of struct taskstsats as follows
 241                  *
 242                  *      per-task-foo(stats, tsk);
 243                  */
 244                 delayacct_add_tsk(stats, tsk);
 245
 246                 stats->nvcsw += tsk->nvcsw;
 247                 stats->nivcsw += tsk->nivcsw;
 248         } while_each_thread(first, tsk);
 249
 250         unlock_task_sighand(first, &flags);
 251         rc = 0;
 252 out:
 253         rcu_read_unlock();
 254
 255         stats->version = TASKSTATS_VERSION;
 256         /*
 257          * Accounting subsystems can also add calls here to modify
 258          * fields of taskstats.
 259          */
 260         return rc;
 261 }
 262
 263 static void fill_tgid_exit(struct task_struct *tsk)
 264 {
 265         unsigned long flags;
 266
 267         spin_lock_irqsave(&tsk->sighand->siglock, flags);
 268         if (!tsk->signal->stats)
 269                 goto ret;
 270
 271         /*
 272          * Each accounting subsystem calls its functions here to
 273          * accumalate its per-task stats for tsk, into the per-tgid structure
 274          *
 275          *      per-task-foo(tsk->signal->stats, tsk);
 276          */
 277         delayacct_add_tsk(tsk->signal->stats, tsk);
 278 ret:
 279         spin_unlock_irqrestore(&tsk->sighand->siglock, flags);
 280         return;
 281 }
 282
 283 static int add_del_listener(pid_t pid, const struct cpumask *mask, int isadd)
 284 {
 285         struct listener_list *listeners;
 286         struct listener *s, *tmp, *s2;
 287         unsigned int cpu;
 288         int ret = 0;
 289
 290         if (!cpumask_subset(mask, cpu_possible_mask))
 291                 return -EINVAL;
 292
 293         if (current_user_ns() != &init_user_ns)
 294                 return -EINVAL;
 295
 296         if (task_active_pid_ns(current) != &init_pid_ns)
 297                 return -EINVAL;
 298
 299         if (isadd == REGISTER) {
 300                 for_each_cpu(cpu, mask) {
 301                         s = kmalloc_node(sizeof(struct listener),
 302                                         GFP_KERNEL, cpu_to_node(cpu));
 303                         if (!s) {
 304                                 ret = -ENOMEM;
 305                                 goto cleanup;
 306                         }
 307                         s->pid = pid;
 308                         s->valid = 1;
 309
 310                         listeners = &per_cpu(listener_array, cpu);
 311                         down_write(&listeners->sem);
 312                         list_for_each_entry(s2, &listeners->list, list) {
 313                                 if (s2->pid == pid && s2->valid)
 314                                         goto exists;
 315                         }
 316                         list_add(&s->list, &listeners->list);
 317                         s = NULL;
 318 exists:
 319                         up_write(&listeners->sem);
 320                         kfree(s); /* nop if NULL */
 321                 }
 322                 return 0;
 323         }
 324
 325         /* Deregister or cleanup */
 326 cleanup:
 327         for_each_cpu(cpu, mask) {
 328                 listeners = &per_cpu(listener_array, cpu);
 329                 down_write(&listeners->sem);
 330                 list_for_each_entry_safe(s, tmp, &listeners->list, list) {
 331                         if (s->pid == pid) {
 332                                 list_del(&s->list);
 333                                 kfree(s);
 334                                 break;
 335                         }
 336                 }
 337                 up_write(&listeners->sem);
 338         }
 339         return ret;
 340 }
 341
 342 static int parse(struct nlattr *na, struct cpumask *mask)
 343 {
 344         char *data;
 345         int len;
 346         int ret;
 347
 348         if (na == NULL)
 349                 return 1;
 350         len = nla_len(na);
 351         if (len > TASKSTATS_CPUMASK_MAXLEN)
 352                 return -E2BIG;
 353         if (len < 1)
 354                 return -EINVAL;
 355         data = kmalloc(len, GFP_KERNEL);
 356         if (!data)
 357                 return -ENOMEM;
 358         nla_strlcpy(data, na, len);
 359         ret = cpulist_parse(data, mask);
 360         kfree(data);
 361         return ret;
 362 }
 363
 364 static struct taskstats *mk_reply(struct sk_buff *skb, int type, u32 pid)
 365 {
 366         struct nlattr *na, *ret;
 367         int aggr;
 368
 369         aggr = (type == TASKSTATS_TYPE_PID)
 370                         ? TASKSTATS_TYPE_AGGR_PID
 371                         : TASKSTATS_TYPE_AGGR_TGID;
 372
 373         na = nla_nest_start(skb, aggr);
 374         if (!na)
 375                 goto err;
 376
 377         if (nla_put(skb, type, sizeof(pid), &pid) < 0) {
 378                 nla_nest_cancel(skb, na);
 379                 goto err;
 380         }
 381         ret = nla_reserve_64bit(skb, TASKSTATS_TYPE_STATS,
 382                                 sizeof(struct taskstats), TASKSTATS_TYPE_NULL);
 383         if (!ret) {
 384                 nla_nest_cancel(skb, na);
 385                 goto err;
 386         }
 387         nla_nest_end(skb, na);
 388
 389         return nla_data(ret);
 390 err:
 391         return NULL;
 392 }
 393
 394 static int cgroupstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
 395 {
 396         int rc = 0;
 397         struct sk_buff *rep_skb;
 398         struct cgroupstats *stats;
 399         struct nlattr *na;
 400         size_t size;
 401         u32 fd;
 402         struct fd f;
 403
 404         na = info->attrs[CGROUPSTATS_CMD_ATTR_FD];
 405         if (!na)
 406                 return -EINVAL;
 407
 408         fd = nla_get_u32(info->attrs[CGROUPSTATS_CMD_ATTR_FD]);
 409         f = fdget(fd);
 410         if (!f.file)
 411                 return 0;
 412
 413         size = nla_total_size(sizeof(struct cgroupstats));
 414
 415         rc = prepare_reply(info, CGROUPSTATS_CMD_NEW, &rep_skb,
 416                                 size);
 417         if (rc < 0)
 418                 goto err;
 419
 420         na = nla_reserve(rep_skb, CGROUPSTATS_TYPE_CGROUP_STATS,
 421                                 sizeof(struct cgroupstats));
 422         if (na == NULL) {
 423                 nlmsg_free(rep_skb);
 424                 rc = -EMSGSIZE;
 425                 goto err;
 426         }
 427
 428         stats = nla_data(na);
 429         memset(stats, 0, sizeof(*stats));
 430
 431         rc = cgroupstats_build(stats, f.file->f_path.dentry);
 432         if (rc < 0) {
 433                 nlmsg_free(rep_skb);
 434                 goto err;
 435         }
 436
 437         rc = send_reply(rep_skb, info);
 438
 439 err:
 440         fdput(f);
 441         return rc;
 442 }
 443
 444 static int cmd_attr_register_cpumask(struct genl_info *info)
 445 {
 446         cpumask_var_t mask;
 447         int rc;
 448
 449         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
 450                 return -ENOMEM;
 451         rc = parse(info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK], mask);
 452         if (rc < 0)
 453                 goto out;
 454         rc = add_del_listener(info->snd_portid, mask, REGISTER);
 455 out:
 456         free_cpumask_var(mask);
 457         return rc;
 458 }
 459
 460 static int cmd_attr_deregister_cpumask(struct genl_info *info)
 461 {
 462         cpumask_var_t mask;
 463         int rc;
 464
 465         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
 466                 return -ENOMEM;
 467         rc = parse(info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK], mask);
 468         if (rc < 0)
 469                 goto out;
 470         rc = add_del_listener(info->snd_portid, mask, DEREGISTER);
 471 out:
 472         free_cpumask_var(mask);
 473         return rc;
 474 }
 475
 476 static size_t taskstats_packet_size(void)
 477 {
 478         size_t size;
 479
 480         size = nla_total_size(sizeof(u32)) +
 481                 nla_total_size_64bit(sizeof(struct taskstats)) +
 482                 nla_total_size(0);
 483
 484         return size;
 485 }
 486
 487 static int cmd_attr_pid(struct genl_info *info)
 488 {
 489         struct taskstats *stats;
 490         struct sk_buff *rep_skb;
 491         size_t size;
 492         u32 pid;
 493         int rc;
 494
 495         size = taskstats_packet_size();
 496
 497         rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
 498         if (rc < 0)
 499                 return rc;
 500
 501         rc = -EINVAL;
 502         pid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_PID]);
 503         stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID, pid);
 504         if (!stats)
 505                 goto err;
 506
 507         rc = fill_stats_for_pid(pid, stats);
 508         if (rc < 0)
 509                 goto err;
 510         return send_reply(rep_skb, info);
 511 err:
 512         nlmsg_free(rep_skb);
 513         return rc;
 514 }
 515
 516 static int cmd_attr_tgid(struct genl_info *info)
 517 {
 518         struct taskstats *stats;
 519         struct sk_buff *rep_skb;
 520         size_t size;
 521         u32 tgid;
 522         int rc;
 523
 524         size = taskstats_packet_size();
 525
 526         rc = prepare_reply(info, TASKSTATS_CMD_NEW, &rep_skb, size);
 527         if (rc < 0)
 528                 return rc;
 529
 530         rc = -EINVAL;
 531         tgid = nla_get_u32(info->attrs[TASKSTATS_CMD_ATTR_TGID]);
 532         stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID, tgid);
 533         if (!stats)
 534                 goto err;
 535
 536         rc = fill_stats_for_tgid(tgid, stats);
 537         if (rc < 0)
 538                 goto err;
 539         return send_reply(rep_skb, info);
 540 err:
 541         nlmsg_free(rep_skb);
 542         return rc;
 543 }
 544
 545 static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
 546 {
 547         if (info->attrs[TASKSTATS_CMD_ATTR_REGISTER_CPUMASK])
 548                 return cmd_attr_register_cpumask(info);
 549         else if (info->attrs[TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK])
 550                 return cmd_attr_deregister_cpumask(info);
 551         else if (info->attrs[TASKSTATS_CMD_ATTR_PID])
 552                 return cmd_attr_pid(info);
 553         else if (info->attrs[TASKSTATS_CMD_ATTR_TGID])
 554                 return cmd_attr_tgid(info);
 555         else
 556                 return -EINVAL;
 557 }
 558
 559 static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
 560 {
 561         struct signal_struct *sig = tsk->signal;
 562         struct taskstats *stats;
 563
 564         if (sig->stats || thread_group_empty(tsk))
 565                 goto ret;
 566
 567         /* No problem if kmem_cache_zalloc() fails */
 568         stats = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
 569
 570         spin_lock_irq(&tsk->sighand->siglock);
 571         if (!sig->stats) {
 572                 sig->stats = stats;
 573                 stats = NULL;
 574         }
 575         spin_unlock_irq(&tsk->sighand->siglock);
 576
 577         if (stats)
 578                 kmem_cache_free(taskstats_cache, stats);
 579 ret:
 580         return sig->stats;
 581 }
 582
 583 /* Send pid data out on exit */
 584 void taskstats_exit(struct task_struct *tsk, int group_dead)
 585 {
 586         int rc;
 587         struct listener_list *listeners;
 588         struct taskstats *stats;
 589         struct sk_buff *rep_skb;
 590         size_t size;
 591         int is_thread_group;
 592
 593         if (!family_registered)
 594                 return;
 595
 596         /*
 597          * Size includes space for nested attributes
 598          */
 599         size = taskstats_packet_size();
 600
 601         is_thread_group = !!taskstats_tgid_alloc(tsk);
 602         if (is_thread_group) {
 603                 /* PID + STATS + TGID + STATS */
 604                 size = 2 * size;
 605                 /* fill the tsk->signal->stats structure */
 606                 fill_tgid_exit(tsk);
 607         }
 608
 609         listeners = raw_cpu_ptr(&listener_array);
 610         if (list_empty(&listeners->list))
 611                 return;
 612
 613         rc = prepare_reply(NULL, TASKSTATS_CMD_NEW, &rep_skb, size);
 614         if (rc < 0)
 615                 return;
 616
 617         stats = mk_reply(rep_skb, TASKSTATS_TYPE_PID,
 618                          task_pid_nr_ns(tsk, &init_pid_ns));
 619         if (!stats)
 620                 goto err;
 621
 622         fill_stats(&init_user_ns, &init_pid_ns, tsk, stats);
 623
 624         /*
 625          * Doesn't matter if tsk is the leader or the last group member leaving
 626          */
 627         if (!is_thread_group || !group_dead)
 628                 goto send;
 629
 630         stats = mk_reply(rep_skb, TASKSTATS_TYPE_TGID,
 631                          task_tgid_nr_ns(tsk, &init_pid_ns));
 632         if (!stats)
 633                 goto err;
 634
 635         memcpy(stats, tsk->signal->stats, sizeof(*stats));
 636
 637 send:
 638         send_cpu_listeners(rep_skb, listeners);
 639         return;
 640 err:
 641         nlmsg_free(rep_skb);
 642 }
 643
 644 static const struct genl_ops taskstats_ops[] = {
 645         {
 646                 .cmd            = TASKSTATS_CMD_GET,
 647                 .doit           = taskstats_user_cmd,
 648                 .policy         = taskstats_cmd_get_policy,
 649                 .flags          = GENL_ADMIN_PERM,
 650         },
 651         {
 652                 .cmd            = CGROUPSTATS_CMD_GET,
 653                 .doit           = cgroupstats_user_cmd,
 654                 .policy         = cgroupstats_cmd_get_policy,
 655         },
 656 };
 657
 658 /* Needed early in initialization */
 659 void __init taskstats_init_early(void)
 660 {
 661         unsigned int i;
 662
 663         taskstats_cache = KMEM_CACHE(taskstats, SLAB_PANIC);
 664         for_each_possible_cpu(i) {
 665                 INIT_LIST_HEAD(&(per_cpu(listener_array, i).list));
 666                 init_rwsem(&(per_cpu(listener_array, i).sem));
 667         }
 668 }
 669
 670 static int __init taskstats_init(void)
 671 {
 672         int rc;
 673
 674         rc = genl_register_family_with_ops(&family, taskstats_ops);
 675         if (rc)
 676                 return rc;
 677
 678         family_registered = 1;
 679         pr_info("registered taskstats version %d\n", TASKSTATS_GENL_VERSION);
 680         return 0;
 681 }
 682
 683 /*
 684  * late initcall ensures initialization of statistics collection
 685  * mechanisms precedes initialization of the taskstats interface
 686  */
 687 late_initcall(taskstats_init);