tools/accounting/getdelays.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* getdelays.c
   3  *
   4  * Utility to get per-pid and per-tgid delay accounting statistics
   5  * Also illustrates usage of the taskstats interface
   6  *
   7  * Copyright (C) Shailabh Nagar, IBM Corp. 2005
   8  * Copyright (C) Balbir Singh, IBM Corp. 2006
   9  * Copyright (c) Jay Lan, SGI. 2006
  10  *
  11  * Compile with
  12  *      gcc -I/usr/src/linux/include getdelays.c -o getdelays
  13  */
  14
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <errno.h>
  18 #include <unistd.h>
  19 #include <poll.h>
  20 #include <string.h>
  21 #include <fcntl.h>
  22 #include <sys/types.h>
  23 #include <sys/stat.h>
  24 #include <sys/socket.h>
  25 #include <sys/wait.h>
  26 #include <signal.h>
  27
  28 #include <linux/genetlink.h>
  29 #include <linux/taskstats.h>
  30 #include <linux/cgroupstats.h>
  31
  32 /*
  33  * Generic macros for dealing with netlink sockets. Might be duplicated
  34  * elsewhere. It is recommended that commercial grade applications use
  35  * libnl or libnetlink and use the interfaces provided by the library
  36  */
  37 #define GENLMSG_DATA(glh)       ((void *)(NLMSG_DATA(glh) + GENL_HDRLEN))
  38 #define GENLMSG_PAYLOAD(glh)    (NLMSG_PAYLOAD(glh, 0) - GENL_HDRLEN)
  39 #define NLA_DATA(na)            ((void *)((char*)(na) + NLA_HDRLEN))
  40 #define NLA_PAYLOAD(len)        (len - NLA_HDRLEN)
  41
  42 #define err(code, fmt, arg...)                  \
  43         do {                                    \
  44                 fprintf(stderr, fmt, ##arg);    \
  45                 exit(code);                     \
  46         } while (0)
  47
  48 int rcvbufsz;
  49 char name[100];
  50 int dbg;
  51 int print_delays;
  52 int print_io_accounting;
  53 int print_task_context_switch_counts;
  54
  55 #define PRINTF(fmt, arg...) {                   \
  56             if (dbg) {                          \
  57                 printf(fmt, ##arg);             \
  58             }                                   \
  59         }
  60
  61 /* Maximum size of response requested or message sent */
  62 #define MAX_MSG_SIZE    1024
  63 /* Maximum number of cpus expected to be specified in a cpumask */
  64 #define MAX_CPUS        32
  65
  66 struct msgtemplate {
  67         struct nlmsghdr n;
  68         struct genlmsghdr g;
  69         char buf[MAX_MSG_SIZE];
  70 };
  71
  72 char cpumask[100+6*MAX_CPUS];
  73
  74 static void usage(void)
  75 {
  76         fprintf(stderr, "getdelays [-dilv] [-w logfile] [-r bufsize] "
  77                         "[-m cpumask] [-t tgid] [-p pid]\n");
  78         fprintf(stderr, "  -d: print delayacct stats\n");
  79         fprintf(stderr, "  -i: print IO accounting (works only with -p)\n");
  80         fprintf(stderr, "  -l: listen forever\n");
  81         fprintf(stderr, "  -v: debug on\n");
  82         fprintf(stderr, "  -C: container path\n");
  83 }
  84
  85 /*
  86  * Create a raw netlink socket and bind
  87  */
  88 static int create_nl_socket(int protocol)
  89 {
  90         int fd;
  91         struct sockaddr_nl local;
  92
  93         fd = socket(AF_NETLINK, SOCK_RAW, protocol);
  94         if (fd < 0)
  95                 return -1;
  96
  97         if (rcvbufsz)
  98                 if (setsockopt(fd, SOL_SOCKET, SO_RCVBUF,
  99                                 &rcvbufsz, sizeof(rcvbufsz)) < 0) {
 100                         fprintf(stderr, "Unable to set socket rcv buf size to %d\n",
 101                                 rcvbufsz);
 102                         goto error;
 103                 }
 104
 105         memset(&local, 0, sizeof(local));
 106         local.nl_family = AF_NETLINK;
 107
 108         if (bind(fd, (struct sockaddr *) &local, sizeof(local)) < 0)
 109                 goto error;
 110
 111         return fd;
 112 error:
 113         close(fd);
 114         return -1;
 115 }
 116
 117
 118 static int send_cmd(int sd, __u16 nlmsg_type, __u32 nlmsg_pid,
 119              __u8 genl_cmd, __u16 nla_type,
 120              void *nla_data, int nla_len)
 121 {
 122         struct nlattr *na;
 123         struct sockaddr_nl nladdr;
 124         int r, buflen;
 125         char *buf;
 126
 127         struct msgtemplate msg;
 128
 129         msg.n.nlmsg_len = NLMSG_LENGTH(GENL_HDRLEN);
 130         msg.n.nlmsg_type = nlmsg_type;
 131         msg.n.nlmsg_flags = NLM_F_REQUEST;
 132         msg.n.nlmsg_seq = 0;
 133         msg.n.nlmsg_pid = nlmsg_pid;
 134         msg.g.cmd = genl_cmd;
 135         msg.g.version = 0x1;
 136         na = (struct nlattr *) GENLMSG_DATA(&msg);
 137         na->nla_type = nla_type;
 138         na->nla_len = nla_len + NLA_HDRLEN;
 139         memcpy(NLA_DATA(na), nla_data, nla_len);
 140         msg.n.nlmsg_len += NLMSG_ALIGN(na->nla_len);
 141
 142         buf = (char *) &msg;
 143         buflen = msg.n.nlmsg_len ;
 144         memset(&nladdr, 0, sizeof(nladdr));
 145         nladdr.nl_family = AF_NETLINK;
 146         while ((r = sendto(sd, buf, buflen, 0, (struct sockaddr *) &nladdr,
 147                            sizeof(nladdr))) < buflen) {
 148                 if (r > 0) {
 149                         buf += r;
 150                         buflen -= r;
 151                 } else if (errno != EAGAIN)
 152                         return -1;
 153         }
 154         return 0;
 155 }
 156
 157
 158 /*
 159  * Probe the controller in genetlink to find the family id
 160  * for the TASKSTATS family
 161  */
 162 static int get_family_id(int sd)
 163 {
 164         struct {
 165                 struct nlmsghdr n;
 166                 struct genlmsghdr g;
 167                 char buf[256];
 168         } ans;
 169
 170         int id = 0, rc;
 171         struct nlattr *na;
 172         int rep_len;
 173
 174         strcpy(name, TASKSTATS_GENL_NAME);
 175         rc = send_cmd(sd, GENL_ID_CTRL, getpid(), CTRL_CMD_GETFAMILY,
 176                         CTRL_ATTR_FAMILY_NAME, (void *)name,
 177                         strlen(TASKSTATS_GENL_NAME)+1);
 178         if (rc < 0)
 179                 return 0;       /* sendto() failure? */
 180
 181         rep_len = recv(sd, &ans, sizeof(ans), 0);
 182         if (ans.n.nlmsg_type == NLMSG_ERROR ||
 183             (rep_len < 0) || !NLMSG_OK((&ans.n), rep_len))
 184                 return 0;
 185
 186         na = (struct nlattr *) GENLMSG_DATA(&ans);
 187         na = (struct nlattr *) ((char *) na + NLA_ALIGN(na->nla_len));
 188         if (na->nla_type == CTRL_ATTR_FAMILY_ID) {
 189                 id = *(__u16 *) NLA_DATA(na);
 190         }
 191         return id;
 192 }
 193
 194 #define average_ms(t, c) (t / 1000000ULL / (c ? c : 1))
 195
 196 static void print_delayacct(struct taskstats *t)
 197 {
 198         printf("\n\nCPU   %15s%15s%15s%15s%15s\n"
 199                "      %15llu%15llu%15llu%15llu%15.3fms\n"
 200                "IO    %15s%15s%15s\n"
 201                "      %15llu%15llu%15llums\n"
 202                "SWAP  %15s%15s%15s\n"
 203                "      %15llu%15llu%15llums\n"
 204                "RECLAIM  %12s%15s%15s\n"
 205                "      %15llu%15llu%15llums\n"
 206                "THRASHING%12s%15s%15s\n"
 207                "      %15llu%15llu%15llums\n"
 208                "COMPACT  %12s%15s%15s\n"
 209                "      %15llu%15llu%15llums\n"
 210                "WPCOPY   %12s%15s%15s\n"
 211                "      %15llu%15llu%15llums\n",
 212                "count", "real total", "virtual total",
 213                "delay total", "delay average",
 214                (unsigned long long)t->cpu_count,
 215                (unsigned long long)t->cpu_run_real_total,
 216                (unsigned long long)t->cpu_run_virtual_total,
 217                (unsigned long long)t->cpu_delay_total,
 218                average_ms((double)t->cpu_delay_total, t->cpu_count),
 219                "count", "delay total", "delay average",
 220                (unsigned long long)t->blkio_count,
 221                (unsigned long long)t->blkio_delay_total,
 222                average_ms(t->blkio_delay_total, t->blkio_count),
 223                "count", "delay total", "delay average",
 224                (unsigned long long)t->swapin_count,
 225                (unsigned long long)t->swapin_delay_total,
 226                average_ms(t->swapin_delay_total, t->swapin_count),
 227                "count", "delay total", "delay average",
 228                (unsigned long long)t->freepages_count,
 229                (unsigned long long)t->freepages_delay_total,
 230                average_ms(t->freepages_delay_total, t->freepages_count),
 231                "count", "delay total", "delay average",
 232                (unsigned long long)t->thrashing_count,
 233                (unsigned long long)t->thrashing_delay_total,
 234                average_ms(t->thrashing_delay_total, t->thrashing_count),
 235                "count", "delay total", "delay average",
 236                (unsigned long long)t->compact_count,
 237                (unsigned long long)t->compact_delay_total,
 238                average_ms(t->compact_delay_total, t->compact_count),
 239                "count", "delay total", "delay average",
 240                (unsigned long long)t->wpcopy_count,
 241                (unsigned long long)t->wpcopy_delay_total,
 242                average_ms(t->wpcopy_delay_total, t->wpcopy_count));
 243 }
 244
 245 static void task_context_switch_counts(struct taskstats *t)
 246 {
 247         printf("\n\nTask   %15s%15s\n"
 248                "       %15llu%15llu\n",
 249                "voluntary", "nonvoluntary",
 250                (unsigned long long)t->nvcsw, (unsigned long long)t->nivcsw);
 251 }
 252
 253 static void print_cgroupstats(struct cgroupstats *c)
 254 {
 255         printf("sleeping %llu, blocked %llu, running %llu, stopped %llu, "
 256                 "uninterruptible %llu\n", (unsigned long long)c->nr_sleeping,
 257                 (unsigned long long)c->nr_io_wait,
 258                 (unsigned long long)c->nr_running,
 259                 (unsigned long long)c->nr_stopped,
 260                 (unsigned long long)c->nr_uninterruptible);
 261 }
 262
 263
 264 static void print_ioacct(struct taskstats *t)
 265 {
 266         printf("%s: read=%llu, write=%llu, cancelled_write=%llu\n",
 267                 t->ac_comm,
 268                 (unsigned long long)t->read_bytes,
 269                 (unsigned long long)t->write_bytes,
 270                 (unsigned long long)t->cancelled_write_bytes);
 271 }
 272
 273 int main(int argc, char *argv[])
 274 {
 275         int c, rc, rep_len, aggr_len, len2;
 276         int cmd_type = TASKSTATS_CMD_ATTR_UNSPEC;
 277         __u16 id;
 278         __u32 mypid;
 279
 280         struct nlattr *na;
 281         int nl_sd = -1;
 282         int len = 0;
 283         pid_t tid = 0;
 284         pid_t rtid = 0;
 285
 286         int fd = 0;
 287         int write_file = 0;
 288         int maskset = 0;
 289         char *logfile = NULL;
 290         int loop = 0;
 291         int containerset = 0;
 292         char *containerpath = NULL;
 293         int cfd = 0;
 294         int forking = 0;
 295         sigset_t sigset;
 296
 297         struct msgtemplate msg;
 298
 299         while (!forking) {
 300                 c = getopt(argc, argv, "qdiw:r:m:t:p:vlC:c:");
 301                 if (c < 0)
 302                         break;
 303
 304                 switch (c) {
 305                 case 'd':
 306                         printf("print delayacct stats ON\n");
 307                         print_delays = 1;
 308                         break;
 309                 case 'i':
 310                         printf("printing IO accounting\n");
 311                         print_io_accounting = 1;
 312                         break;
 313                 case 'q':
 314                         printf("printing task/process context switch rates\n");
 315                         print_task_context_switch_counts = 1;
 316                         break;
 317                 case 'C':
 318                         containerset = 1;
 319                         containerpath = optarg;
 320                         break;
 321                 case 'w':
 322                         logfile = strdup(optarg);
 323                         printf("write to file %s\n", logfile);
 324                         write_file = 1;
 325                         break;
 326                 case 'r':
 327                         rcvbufsz = atoi(optarg);
 328                         printf("receive buf size %d\n", rcvbufsz);
 329                         if (rcvbufsz < 0)
 330                                 err(1, "Invalid rcv buf size\n");
 331                         break;
 332                 case 'm':
 333                         strncpy(cpumask, optarg, sizeof(cpumask));
 334                         cpumask[sizeof(cpumask) - 1] = '\0';
 335                         maskset = 1;
 336                         printf("cpumask %s maskset %d\n", cpumask, maskset);
 337                         break;
 338                 case 't':
 339                         tid = atoi(optarg);
 340                         if (!tid)
 341                                 err(1, "Invalid tgid\n");
 342                         cmd_type = TASKSTATS_CMD_ATTR_TGID;
 343                         break;
 344                 case 'p':
 345                         tid = atoi(optarg);
 346                         if (!tid)
 347                                 err(1, "Invalid pid\n");
 348                         cmd_type = TASKSTATS_CMD_ATTR_PID;
 349                         break;
 350                 case 'c':
 351
 352                         /* Block SIGCHLD for sigwait() later */
 353                         if (sigemptyset(&sigset) == -1)
 354                                 err(1, "Failed to empty sigset");
 355                         if (sigaddset(&sigset, SIGCHLD))
 356                                 err(1, "Failed to set sigchld in sigset");
 357                         sigprocmask(SIG_BLOCK, &sigset, NULL);
 358
 359                         /* fork/exec a child */
 360                         tid = fork();
 361                         if (tid < 0)
 362                                 err(1, "Fork failed\n");
 363                         if (tid == 0)
 364                                 if (execvp(argv[optind - 1],
 365                                     &argv[optind - 1]) < 0)
 366                                         exit(-1);
 367
 368                         /* Set the command type and avoid further processing */
 369                         cmd_type = TASKSTATS_CMD_ATTR_PID;
 370                         forking = 1;
 371                         break;
 372                 case 'v':
 373                         printf("debug on\n");
 374                         dbg = 1;
 375                         break;
 376                 case 'l':
 377                         printf("listen forever\n");
 378                         loop = 1;
 379                         break;
 380                 default:
 381                         usage();
 382                         exit(-1);
 383                 }
 384         }
 385
 386         if (write_file) {
 387                 fd = open(logfile, O_WRONLY | O_CREAT | O_TRUNC,
 388                           S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
 389                 if (fd == -1) {
 390                         perror("Cannot open output file\n");
 391                         exit(1);
 392                 }
 393         }
 394
 395         nl_sd = create_nl_socket(NETLINK_GENERIC);
 396         if (nl_sd < 0)
 397                 err(1, "error creating Netlink socket\n");
 398
 399
 400         mypid = getpid();
 401         id = get_family_id(nl_sd);
 402         if (!id) {
 403                 fprintf(stderr, "Error getting family id, errno %d\n", errno);
 404                 goto err;
 405         }
 406         PRINTF("family id %d\n", id);
 407
 408         if (maskset) {
 409                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 410                               TASKSTATS_CMD_ATTR_REGISTER_CPUMASK,
 411                               &cpumask, strlen(cpumask) + 1);
 412                 PRINTF("Sent register cpumask, retval %d\n", rc);
 413                 if (rc < 0) {
 414                         fprintf(stderr, "error sending register cpumask\n");
 415                         goto err;
 416                 }
 417         }
 418
 419         if (tid && containerset) {
 420                 fprintf(stderr, "Select either -t or -C, not both\n");
 421                 goto err;
 422         }
 423
 424         /*
 425          * If we forked a child, wait for it to exit. Cannot use waitpid()
 426          * as all the delicious data would be reaped as part of the wait
 427          */
 428         if (tid && forking) {
 429                 int sig_received;
 430                 sigwait(&sigset, &sig_received);
 431         }
 432
 433         if (tid) {
 434                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 435                               cmd_type, &tid, sizeof(__u32));
 436                 PRINTF("Sent pid/tgid, retval %d\n", rc);
 437                 if (rc < 0) {
 438                         fprintf(stderr, "error sending tid/tgid cmd\n");
 439                         goto done;
 440                 }
 441         }
 442
 443         if (containerset) {
 444                 cfd = open(containerpath, O_RDONLY);
 445                 if (cfd < 0) {
 446                         perror("error opening container file");
 447                         goto err;
 448                 }
 449                 rc = send_cmd(nl_sd, id, mypid, CGROUPSTATS_CMD_GET,
 450                               CGROUPSTATS_CMD_ATTR_FD, &cfd, sizeof(__u32));
 451                 if (rc < 0) {
 452                         perror("error sending cgroupstats command");
 453                         goto err;
 454                 }
 455         }
 456         if (!maskset && !tid && !containerset) {
 457                 usage();
 458                 goto err;
 459         }
 460
 461         do {
 462                 rep_len = recv(nl_sd, &msg, sizeof(msg), 0);
 463                 PRINTF("received %d bytes\n", rep_len);
 464
 465                 if (rep_len < 0) {
 466                         fprintf(stderr, "nonfatal reply error: errno %d\n",
 467                                 errno);
 468                         continue;
 469                 }
 470                 if (msg.n.nlmsg_type == NLMSG_ERROR ||
 471                     !NLMSG_OK((&msg.n), rep_len)) {
 472                         struct nlmsgerr *err = NLMSG_DATA(&msg);
 473                         fprintf(stderr, "fatal reply error,  errno %d\n",
 474                                 err->error);
 475                         goto done;
 476                 }
 477
 478                 PRINTF("nlmsghdr size=%zu, nlmsg_len=%d, rep_len=%d\n",
 479                        sizeof(struct nlmsghdr), msg.n.nlmsg_len, rep_len);
 480
 481
 482                 rep_len = GENLMSG_PAYLOAD(&msg.n);
 483
 484                 na = (struct nlattr *) GENLMSG_DATA(&msg);
 485                 len = 0;
 486                 while (len < rep_len) {
 487                         len += NLA_ALIGN(na->nla_len);
 488                         switch (na->nla_type) {
 489                         case TASKSTATS_TYPE_AGGR_TGID:
 490                                 /* Fall through */
 491                         case TASKSTATS_TYPE_AGGR_PID:
 492                                 aggr_len = NLA_PAYLOAD(na->nla_len);
 493                                 len2 = 0;
 494                                 /* For nested attributes, na follows */
 495                                 na = (struct nlattr *) NLA_DATA(na);
 496                                 while (len2 < aggr_len) {
 497                                         switch (na->nla_type) {
 498                                         case TASKSTATS_TYPE_PID:
 499                                                 rtid = *(int *) NLA_DATA(na);
 500                                                 if (print_delays)
 501                                                         printf("PID\t%d\n", rtid);
 502                                                 break;
 503                                         case TASKSTATS_TYPE_TGID:
 504                                                 rtid = *(int *) NLA_DATA(na);
 505                                                 if (print_delays)
 506                                                         printf("TGID\t%d\n", rtid);
 507                                                 break;
 508                                         case TASKSTATS_TYPE_STATS:
 509                                                 if (print_delays)
 510                                                         print_delayacct((struct taskstats *) NLA_DATA(na));
 511                                                 if (print_io_accounting)
 512                                                         print_ioacct((struct taskstats *) NLA_DATA(na));
 513                                                 if (print_task_context_switch_counts)
 514                                                         task_context_switch_counts((struct taskstats *) NLA_DATA(na));
 515                                                 if (fd) {
 516                                                         if (write(fd, NLA_DATA(na), na->nla_len) < 0) {
 517                                                                 err(1,"write error\n");
 518                                                         }
 519                                                 }
 520                                                 if (!loop)
 521                                                         goto done;
 522                                                 break;
 523                                         case TASKSTATS_TYPE_NULL:
 524                                                 break;
 525                                         default:
 526                                                 fprintf(stderr, "Unknown nested"
 527                                                         " nla_type %d\n",
 528                                                         na->nla_type);
 529                                                 break;
 530                                         }
 531                                         len2 += NLA_ALIGN(na->nla_len);
 532                                         na = (struct nlattr *)((char *)na +
 533                                                                NLA_ALIGN(na->nla_len));
 534                                 }
 535                                 break;
 536
 537                         case CGROUPSTATS_TYPE_CGROUP_STATS:
 538                                 print_cgroupstats(NLA_DATA(na));
 539                                 break;
 540                         default:
 541                                 fprintf(stderr, "Unknown nla_type %d\n",
 542                                         na->nla_type);
 543                         case TASKSTATS_TYPE_NULL:
 544                                 break;
 545                         }
 546                         na = (struct nlattr *) (GENLMSG_DATA(&msg) + len);
 547                 }
 548         } while (loop);
 549 done:
 550         if (maskset) {
 551                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 552                               TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK,
 553                               &cpumask, strlen(cpumask) + 1);
 554                 printf("Sent deregister mask, retval %d\n", rc);
 555                 if (rc < 0)
 556                         err(rc, "error sending deregister cpumask\n");
 557         }
 558 err:
 559         close(nl_sd);
 560         if (fd)
 561                 close(fd);
 562         if (cfd)
 563                 close(cfd);
 564         return 0;
 565 }