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