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                "WPCOPY   %12s%15s%15s\n"
 212                "      %15llu%15llu%15llums\n",
 213                "count", "real total", "virtual total",
 214                "delay total", "delay average",
 215                (unsigned long long)t->cpu_count,
 216                (unsigned long long)t->cpu_run_real_total,
 217                (unsigned long long)t->cpu_run_virtual_total,
 218                (unsigned long long)t->cpu_delay_total,
 219                average_ms((double)t->cpu_delay_total, t->cpu_count),
 220                "count", "delay total", "delay average",
 221                (unsigned long long)t->blkio_count,
 222                (unsigned long long)t->blkio_delay_total,
 223                average_ms(t->blkio_delay_total, t->blkio_count),
 224                "count", "delay total", "delay average",
 225                (unsigned long long)t->swapin_count,
 226                (unsigned long long)t->swapin_delay_total,
 227                average_ms(t->swapin_delay_total, t->swapin_count),
 228                "count", "delay total", "delay average",
 229                (unsigned long long)t->freepages_count,
 230                (unsigned long long)t->freepages_delay_total,
 231                average_ms(t->freepages_delay_total, t->freepages_count),
 232                "count", "delay total", "delay average",
 233                (unsigned long long)t->thrashing_count,
 234                (unsigned long long)t->thrashing_delay_total,
 235                average_ms(t->thrashing_delay_total, t->thrashing_count),
 236                "count", "delay total", "delay average",
 237                (unsigned long long)t->compact_count,
 238                (unsigned long long)t->compact_delay_total,
 239                average_ms(t->compact_delay_total, t->compact_count),
 240                "count", "delay total", "delay average",
 241                (unsigned long long)t->wpcopy_count,
 242                (unsigned long long)t->wpcopy_delay_total,
 243                average_ms(t->wpcopy_delay_total, t->wpcopy_count));
 244 }
 245
 246 static void task_context_switch_counts(struct taskstats *t)
 247 {
 248         printf("\n\nTask   %15s%15s\n"
 249                "       %15llu%15llu\n",
 250                "voluntary", "nonvoluntary",
 251                (unsigned long long)t->nvcsw, (unsigned long long)t->nivcsw);
 252 }
 253
 254 static void print_cgroupstats(struct cgroupstats *c)
 255 {
 256         printf("sleeping %llu, blocked %llu, running %llu, stopped %llu, "
 257                 "uninterruptible %llu\n", (unsigned long long)c->nr_sleeping,
 258                 (unsigned long long)c->nr_io_wait,
 259                 (unsigned long long)c->nr_running,
 260                 (unsigned long long)c->nr_stopped,
 261                 (unsigned long long)c->nr_uninterruptible);
 262 }
 263
 264
 265 static void print_ioacct(struct taskstats *t)
 266 {
 267         printf("%s: read=%llu, write=%llu, cancelled_write=%llu\n",
 268                 t->ac_comm,
 269                 (unsigned long long)t->read_bytes,
 270                 (unsigned long long)t->write_bytes,
 271                 (unsigned long long)t->cancelled_write_bytes);
 272 }
 273
 274 int main(int argc, char *argv[])
 275 {
 276         int c, rc, rep_len, aggr_len, len2;
 277         int cmd_type = TASKSTATS_CMD_ATTR_UNSPEC;
 278         __u16 id;
 279         __u32 mypid;
 280
 281         struct nlattr *na;
 282         int nl_sd = -1;
 283         int len = 0;
 284         pid_t tid = 0;
 285         pid_t rtid = 0;
 286
 287         int fd = 0;
 288         int count = 0;
 289         int write_file = 0;
 290         int maskset = 0;
 291         char *logfile = NULL;
 292         int loop = 0;
 293         int containerset = 0;
 294         char *containerpath = NULL;
 295         int cfd = 0;
 296         int forking = 0;
 297         sigset_t sigset;
 298
 299         struct msgtemplate msg;
 300
 301         while (!forking) {
 302                 c = getopt(argc, argv, "qdiw:r:m:t:p:vlC:c:");
 303                 if (c < 0)
 304                         break;
 305
 306                 switch (c) {
 307                 case 'd':
 308                         printf("print delayacct stats ON\n");
 309                         print_delays = 1;
 310                         break;
 311                 case 'i':
 312                         printf("printing IO accounting\n");
 313                         print_io_accounting = 1;
 314                         break;
 315                 case 'q':
 316                         printf("printing task/process context switch rates\n");
 317                         print_task_context_switch_counts = 1;
 318                         break;
 319                 case 'C':
 320                         containerset = 1;
 321                         containerpath = optarg;
 322                         break;
 323                 case 'w':
 324                         logfile = strdup(optarg);
 325                         printf("write to file %s\n", logfile);
 326                         write_file = 1;
 327                         break;
 328                 case 'r':
 329                         rcvbufsz = atoi(optarg);
 330                         printf("receive buf size %d\n", rcvbufsz);
 331                         if (rcvbufsz < 0)
 332                                 err(1, "Invalid rcv buf size\n");
 333                         break;
 334                 case 'm':
 335                         strncpy(cpumask, optarg, sizeof(cpumask));
 336                         cpumask[sizeof(cpumask) - 1] = '\0';
 337                         maskset = 1;
 338                         printf("cpumask %s maskset %d\n", cpumask, maskset);
 339                         break;
 340                 case 't':
 341                         tid = atoi(optarg);
 342                         if (!tid)
 343                                 err(1, "Invalid tgid\n");
 344                         cmd_type = TASKSTATS_CMD_ATTR_TGID;
 345                         break;
 346                 case 'p':
 347                         tid = atoi(optarg);
 348                         if (!tid)
 349                                 err(1, "Invalid pid\n");
 350                         cmd_type = TASKSTATS_CMD_ATTR_PID;
 351                         break;
 352                 case 'c':
 353
 354                         /* Block SIGCHLD for sigwait() later */
 355                         if (sigemptyset(&sigset) == -1)
 356                                 err(1, "Failed to empty sigset");
 357                         if (sigaddset(&sigset, SIGCHLD))
 358                                 err(1, "Failed to set sigchld in sigset");
 359                         sigprocmask(SIG_BLOCK, &sigset, NULL);
 360
 361                         /* fork/exec a child */
 362                         tid = fork();
 363                         if (tid < 0)
 364                                 err(1, "Fork failed\n");
 365                         if (tid == 0)
 366                                 if (execvp(argv[optind - 1],
 367                                     &argv[optind - 1]) < 0)
 368                                         exit(-1);
 369
 370                         /* Set the command type and avoid further processing */
 371                         cmd_type = TASKSTATS_CMD_ATTR_PID;
 372                         forking = 1;
 373                         break;
 374                 case 'v':
 375                         printf("debug on\n");
 376                         dbg = 1;
 377                         break;
 378                 case 'l':
 379                         printf("listen forever\n");
 380                         loop = 1;
 381                         break;
 382                 default:
 383                         usage();
 384                         exit(-1);
 385                 }
 386         }
 387
 388         if (write_file) {
 389                 fd = open(logfile, O_WRONLY | O_CREAT | O_TRUNC,
 390                           S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
 391                 if (fd == -1) {
 392                         perror("Cannot open output file\n");
 393                         exit(1);
 394                 }
 395         }
 396
 397         nl_sd = create_nl_socket(NETLINK_GENERIC);
 398         if (nl_sd < 0)
 399                 err(1, "error creating Netlink socket\n");
 400
 401
 402         mypid = getpid();
 403         id = get_family_id(nl_sd);
 404         if (!id) {
 405                 fprintf(stderr, "Error getting family id, errno %d\n", errno);
 406                 goto err;
 407         }
 408         PRINTF("family id %d\n", id);
 409
 410         if (maskset) {
 411                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 412                               TASKSTATS_CMD_ATTR_REGISTER_CPUMASK,
 413                               &cpumask, strlen(cpumask) + 1);
 414                 PRINTF("Sent register cpumask, retval %d\n", rc);
 415                 if (rc < 0) {
 416                         fprintf(stderr, "error sending register cpumask\n");
 417                         goto err;
 418                 }
 419         }
 420
 421         if (tid && containerset) {
 422                 fprintf(stderr, "Select either -t or -C, not both\n");
 423                 goto err;
 424         }
 425
 426         /*
 427          * If we forked a child, wait for it to exit. Cannot use waitpid()
 428          * as all the delicious data would be reaped as part of the wait
 429          */
 430         if (tid && forking) {
 431                 int sig_received;
 432                 sigwait(&sigset, &sig_received);
 433         }
 434
 435         if (tid) {
 436                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 437                               cmd_type, &tid, sizeof(__u32));
 438                 PRINTF("Sent pid/tgid, retval %d\n", rc);
 439                 if (rc < 0) {
 440                         fprintf(stderr, "error sending tid/tgid cmd\n");
 441                         goto done;
 442                 }
 443         }
 444
 445         if (containerset) {
 446                 cfd = open(containerpath, O_RDONLY);
 447                 if (cfd < 0) {
 448                         perror("error opening container file");
 449                         goto err;
 450                 }
 451                 rc = send_cmd(nl_sd, id, mypid, CGROUPSTATS_CMD_GET,
 452                               CGROUPSTATS_CMD_ATTR_FD, &cfd, sizeof(__u32));
 453                 if (rc < 0) {
 454                         perror("error sending cgroupstats command");
 455                         goto err;
 456                 }
 457         }
 458         if (!maskset && !tid && !containerset) {
 459                 usage();
 460                 goto err;
 461         }
 462
 463         do {
 464                 rep_len = recv(nl_sd, &msg, sizeof(msg), 0);
 465                 PRINTF("received %d bytes\n", rep_len);
 466
 467                 if (rep_len < 0) {
 468                         fprintf(stderr, "nonfatal reply error: errno %d\n",
 469                                 errno);
 470                         continue;
 471                 }
 472                 if (msg.n.nlmsg_type == NLMSG_ERROR ||
 473                     !NLMSG_OK((&msg.n), rep_len)) {
 474                         struct nlmsgerr *err = NLMSG_DATA(&msg);
 475                         fprintf(stderr, "fatal reply error,  errno %d\n",
 476                                 err->error);
 477                         goto done;
 478                 }
 479
 480                 PRINTF("nlmsghdr size=%zu, nlmsg_len=%d, rep_len=%d\n",
 481                        sizeof(struct nlmsghdr), msg.n.nlmsg_len, rep_len);
 482
 483
 484                 rep_len = GENLMSG_PAYLOAD(&msg.n);
 485
 486                 na = (struct nlattr *) GENLMSG_DATA(&msg);
 487                 len = 0;
 488                 while (len < rep_len) {
 489                         len += NLA_ALIGN(na->nla_len);
 490                         switch (na->nla_type) {
 491                         case TASKSTATS_TYPE_AGGR_TGID:
 492                                 /* Fall through */
 493                         case TASKSTATS_TYPE_AGGR_PID:
 494                                 aggr_len = NLA_PAYLOAD(na->nla_len);
 495                                 len2 = 0;
 496                                 /* For nested attributes, na follows */
 497                                 na = (struct nlattr *) NLA_DATA(na);
 498                                 done = 0;
 499                                 while (len2 < aggr_len) {
 500                                         switch (na->nla_type) {
 501                                         case TASKSTATS_TYPE_PID:
 502                                                 rtid = *(int *) NLA_DATA(na);
 503                                                 if (print_delays)
 504                                                         printf("PID\t%d\n", rtid);
 505                                                 break;
 506                                         case TASKSTATS_TYPE_TGID:
 507                                                 rtid = *(int *) NLA_DATA(na);
 508                                                 if (print_delays)
 509                                                         printf("TGID\t%d\n", rtid);
 510                                                 break;
 511                                         case TASKSTATS_TYPE_STATS:
 512                                                 count++;
 513                                                 if (print_delays)
 514                                                         print_delayacct((struct taskstats *) NLA_DATA(na));
 515                                                 if (print_io_accounting)
 516                                                         print_ioacct((struct taskstats *) NLA_DATA(na));
 517                                                 if (print_task_context_switch_counts)
 518                                                         task_context_switch_counts((struct taskstats *) NLA_DATA(na));
 519                                                 if (fd) {
 520                                                         if (write(fd, NLA_DATA(na), na->nla_len) < 0) {
 521                                                                 err(1,"write error\n");
 522                                                         }
 523                                                 }
 524                                                 if (!loop)
 525                                                         goto done;
 526                                                 break;
 527                                         case TASKSTATS_TYPE_NULL:
 528                                                 break;
 529                                         default:
 530                                                 fprintf(stderr, "Unknown nested"
 531                                                         " nla_type %d\n",
 532                                                         na->nla_type);
 533                                                 break;
 534                                         }
 535                                         len2 += NLA_ALIGN(na->nla_len);
 536                                         na = (struct nlattr *)((char *)na +
 537                                                                NLA_ALIGN(na->nla_len));
 538                                 }
 539                                 break;
 540
 541                         case CGROUPSTATS_TYPE_CGROUP_STATS:
 542                                 print_cgroupstats(NLA_DATA(na));
 543                                 break;
 544                         default:
 545                                 fprintf(stderr, "Unknown nla_type %d\n",
 546                                         na->nla_type);
 547                         case TASKSTATS_TYPE_NULL:
 548                                 break;
 549                         }
 550                         na = (struct nlattr *) (GENLMSG_DATA(&msg) + len);
 551                 }
 552         } while (loop);
 553 done:
 554         if (maskset) {
 555                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 556                               TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK,
 557                               &cpumask, strlen(cpumask) + 1);
 558                 printf("Sent deregister mask, retval %d\n", rc);
 559                 if (rc < 0)
 560                         err(rc, "error sending deregister cpumask\n");
 561         }
 562 err:
 563         close(nl_sd);
 564         if (fd)
 565                 close(fd);
 566         if (cfd)
 567                 close(cfd);
 568         return 0;
 569 }