Documentation/accounting/getdelays.c

   1 /* getdelays.c
   2  *
   3  * Utility to get per-pid and per-tgid delay accounting statistics
   4  * Also illustrates usage of the taskstats interface
   5  *
   6  * Copyright (C) Shailabh Nagar, IBM Corp. 2005
   7  * Copyright (C) Balbir Singh, IBM Corp. 2006
   8  * Copyright (c) Jay Lan, SGI. 2006
   9  *
  10  * Compile with
  11  *      gcc -I/usr/src/linux/include getdelays.c -o getdelays
  12  */
  13
  14 #include <stdio.h>
  15 #include <stdlib.h>
  16 #include <errno.h>
  17 #include <unistd.h>
  18 #include <poll.h>
  19 #include <string.h>
  20 #include <fcntl.h>
  21 #include <sys/types.h>
  22 #include <sys/stat.h>
  23 #include <sys/socket.h>
  24 #include <sys/wait.h>
  25 #include <signal.h>
  26
  27 #include <linux/genetlink.h>
  28 #include <linux/taskstats.h>
  29 #include <linux/cgroupstats.h>
  30
  31 /*
  32  * Generic macros for dealing with netlink sockets. Might be duplicated
  33  * elsewhere. It is recommended that commercial grade applications use
  34  * libnl or libnetlink and use the interfaces provided by the library
  35  */
  36 #define GENLMSG_DATA(glh)       ((void *)(NLMSG_DATA(glh) + GENL_HDRLEN))
  37 #define GENLMSG_PAYLOAD(glh)    (NLMSG_PAYLOAD(glh, 0) - GENL_HDRLEN)
  38 #define NLA_DATA(na)            ((void *)((char*)(na) + NLA_HDRLEN))
  39 #define NLA_PAYLOAD(len)        (len - NLA_HDRLEN)
  40
  41 #define err(code, fmt, arg...)                  \
  42         do {                                    \
  43                 fprintf(stderr, fmt, ##arg);    \
  44                 exit(code);                     \
  45         } while (0)
  46
  47 int done;
  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 __u64 stime, utime;
  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 "
 102                                         "to %d\n",
 103                                 rcvbufsz);
 104                         return -1;
 105                 }
 106
 107         memset(&local, 0, sizeof(local));
 108         local.nl_family = AF_NETLINK;
 109
 110         if (bind(fd, (struct sockaddr *) &local, sizeof(local)) < 0)
 111                 goto error;
 112
 113         return fd;
 114 error:
 115         close(fd);
 116         return -1;
 117 }
 118
 119
 120 static int send_cmd(int sd, __u16 nlmsg_type, __u32 nlmsg_pid,
 121              __u8 genl_cmd, __u16 nla_type,
 122              void *nla_data, int nla_len)
 123 {
 124         struct nlattr *na;
 125         struct sockaddr_nl nladdr;
 126         int r, buflen;
 127         char *buf;
 128
 129         struct msgtemplate msg;
 130
 131         msg.n.nlmsg_len = NLMSG_LENGTH(GENL_HDRLEN);
 132         msg.n.nlmsg_type = nlmsg_type;
 133         msg.n.nlmsg_flags = NLM_F_REQUEST;
 134         msg.n.nlmsg_seq = 0;
 135         msg.n.nlmsg_pid = nlmsg_pid;
 136         msg.g.cmd = genl_cmd;
 137         msg.g.version = 0x1;
 138         na = (struct nlattr *) GENLMSG_DATA(&msg);
 139         na->nla_type = nla_type;
 140         na->nla_len = nla_len + 1 + NLA_HDRLEN;
 141         memcpy(NLA_DATA(na), nla_data, nla_len);
 142         msg.n.nlmsg_len += NLMSG_ALIGN(na->nla_len);
 143
 144         buf = (char *) &msg;
 145         buflen = msg.n.nlmsg_len ;
 146         memset(&nladdr, 0, sizeof(nladdr));
 147         nladdr.nl_family = AF_NETLINK;
 148         while ((r = sendto(sd, buf, buflen, 0, (struct sockaddr *) &nladdr,
 149                            sizeof(nladdr))) < buflen) {
 150                 if (r > 0) {
 151                         buf += r;
 152                         buflen -= r;
 153                 } else if (errno != EAGAIN)
 154                         return -1;
 155         }
 156         return 0;
 157 }
 158
 159
 160 /*
 161  * Probe the controller in genetlink to find the family id
 162  * for the TASKSTATS family
 163  */
 164 static int get_family_id(int sd)
 165 {
 166         struct {
 167                 struct nlmsghdr n;
 168                 struct genlmsghdr g;
 169                 char buf[256];
 170         } ans;
 171
 172         int id = 0, rc;
 173         struct nlattr *na;
 174         int rep_len;
 175
 176         strcpy(name, TASKSTATS_GENL_NAME);
 177         rc = send_cmd(sd, GENL_ID_CTRL, getpid(), CTRL_CMD_GETFAMILY,
 178                         CTRL_ATTR_FAMILY_NAME, (void *)name,
 179                         strlen(TASKSTATS_GENL_NAME)+1);
 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 static void print_delayacct(struct taskstats *t)
 195 {
 196         printf("\n\nCPU   %15s%15s%15s%15s\n"
 197                "      %15llu%15llu%15llu%15llu\n"
 198                "IO    %15s%15s\n"
 199                "      %15llu%15llu\n"
 200                "SWAP  %15s%15s\n"
 201                "      %15llu%15llu\n"
 202                "RECLAIM  %12s%15s\n"
 203                "      %15llu%15llu\n",
 204                "count", "real total", "virtual total", "delay total",
 205                (unsigned long long)t->cpu_count,
 206                (unsigned long long)t->cpu_run_real_total,
 207                (unsigned long long)t->cpu_run_virtual_total,
 208                (unsigned long long)t->cpu_delay_total,
 209                "count", "delay total",
 210                (unsigned long long)t->blkio_count,
 211                (unsigned long long)t->blkio_delay_total,
 212                "count", "delay total",
 213                (unsigned long long)t->swapin_count,
 214                (unsigned long long)t->swapin_delay_total,
 215                "count", "delay total",
 216                (unsigned long long)t->freepages_count,
 217                (unsigned long long)t->freepages_delay_total);
 218 }
 219
 220 static void task_context_switch_counts(struct taskstats *t)
 221 {
 222         printf("\n\nTask   %15s%15s\n"
 223                "       %15llu%15llu\n",
 224                "voluntary", "nonvoluntary",
 225                (unsigned long long)t->nvcsw, (unsigned long long)t->nivcsw);
 226 }
 227
 228 static void print_cgroupstats(struct cgroupstats *c)
 229 {
 230         printf("sleeping %llu, blocked %llu, running %llu, stopped %llu, "
 231                 "uninterruptible %llu\n", (unsigned long long)c->nr_sleeping,
 232                 (unsigned long long)c->nr_io_wait,
 233                 (unsigned long long)c->nr_running,
 234                 (unsigned long long)c->nr_stopped,
 235                 (unsigned long long)c->nr_uninterruptible);
 236 }
 237
 238
 239 static void print_ioacct(struct taskstats *t)
 240 {
 241         printf("%s: read=%llu, write=%llu, cancelled_write=%llu\n",
 242                 t->ac_comm,
 243                 (unsigned long long)t->read_bytes,
 244                 (unsigned long long)t->write_bytes,
 245                 (unsigned long long)t->cancelled_write_bytes);
 246 }
 247
 248 int main(int argc, char *argv[])
 249 {
 250         int c, rc, rep_len, aggr_len, len2;
 251         int cmd_type = TASKSTATS_CMD_ATTR_UNSPEC;
 252         __u16 id;
 253         __u32 mypid;
 254
 255         struct nlattr *na;
 256         int nl_sd = -1;
 257         int len = 0;
 258         pid_t tid = 0;
 259         pid_t rtid = 0;
 260
 261         int fd = 0;
 262         int count = 0;
 263         int write_file = 0;
 264         int maskset = 0;
 265         char *logfile = NULL;
 266         int loop = 0;
 267         int containerset = 0;
 268         char containerpath[1024];
 269         int cfd = 0;
 270         int forking = 0;
 271         sigset_t sigset;
 272
 273         struct msgtemplate msg;
 274
 275         while (!forking) {
 276                 c = getopt(argc, argv, "qdiw:r:m:t:p:vlC:c:");
 277                 if (c < 0)
 278                         break;
 279
 280                 switch (c) {
 281                 case 'd':
 282                         printf("print delayacct stats ON\n");
 283                         print_delays = 1;
 284                         break;
 285                 case 'i':
 286                         printf("printing IO accounting\n");
 287                         print_io_accounting = 1;
 288                         break;
 289                 case 'q':
 290                         printf("printing task/process context switch rates\n");
 291                         print_task_context_switch_counts = 1;
 292                         break;
 293                 case 'C':
 294                         containerset = 1;
 295                         strncpy(containerpath, optarg, strlen(optarg) + 1);
 296                         break;
 297                 case 'w':
 298                         logfile = strdup(optarg);
 299                         printf("write to file %s\n", logfile);
 300                         write_file = 1;
 301                         break;
 302                 case 'r':
 303                         rcvbufsz = atoi(optarg);
 304                         printf("receive buf size %d\n", rcvbufsz);
 305                         if (rcvbufsz < 0)
 306                                 err(1, "Invalid rcv buf size\n");
 307                         break;
 308                 case 'm':
 309                         strncpy(cpumask, optarg, sizeof(cpumask));
 310                         maskset = 1;
 311                         printf("cpumask %s maskset %d\n", cpumask, maskset);
 312                         break;
 313                 case 't':
 314                         tid = atoi(optarg);
 315                         if (!tid)
 316                                 err(1, "Invalid tgid\n");
 317                         cmd_type = TASKSTATS_CMD_ATTR_TGID;
 318                         break;
 319                 case 'p':
 320                         tid = atoi(optarg);
 321                         if (!tid)
 322                                 err(1, "Invalid pid\n");
 323                         cmd_type = TASKSTATS_CMD_ATTR_PID;
 324                         break;
 325                 case 'c':
 326
 327                         /* Block SIGCHLD for sigwait() later */
 328                         if (sigemptyset(&sigset) == -1)
 329                                 err(1, "Failed to empty sigset");
 330                         if (sigaddset(&sigset, SIGCHLD))
 331                                 err(1, "Failed to set sigchld in sigset");
 332                         sigprocmask(SIG_BLOCK, &sigset, NULL);
 333
 334                         /* fork/exec a child */
 335                         tid = fork();
 336                         if (tid < 0)
 337                                 err(1, "Fork failed\n");
 338                         if (tid == 0)
 339                                 if (execvp(argv[optind - 1],
 340                                     &argv[optind - 1]) < 0)
 341                                         exit(-1);
 342
 343                         /* Set the command type and avoid further processing */
 344                         cmd_type = TASKSTATS_CMD_ATTR_PID;
 345                         forking = 1;
 346                         break;
 347                 case 'v':
 348                         printf("debug on\n");
 349                         dbg = 1;
 350                         break;
 351                 case 'l':
 352                         printf("listen forever\n");
 353                         loop = 1;
 354                         break;
 355                 default:
 356                         usage();
 357                         exit(-1);
 358                 }
 359         }
 360
 361         if (write_file) {
 362                 fd = open(logfile, O_WRONLY | O_CREAT | O_TRUNC,
 363                           S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
 364                 if (fd == -1) {
 365                         perror("Cannot open output file\n");
 366                         exit(1);
 367                 }
 368         }
 369
 370         if ((nl_sd = create_nl_socket(NETLINK_GENERIC)) < 0)
 371                 err(1, "error creating Netlink socket\n");
 372
 373
 374         mypid = getpid();
 375         id = get_family_id(nl_sd);
 376         if (!id) {
 377                 fprintf(stderr, "Error getting family id, errno %d\n", errno);
 378                 goto err;
 379         }
 380         PRINTF("family id %d\n", id);
 381
 382         if (maskset) {
 383                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 384                               TASKSTATS_CMD_ATTR_REGISTER_CPUMASK,
 385                               &cpumask, strlen(cpumask) + 1);
 386                 PRINTF("Sent register cpumask, retval %d\n", rc);
 387                 if (rc < 0) {
 388                         fprintf(stderr, "error sending register cpumask\n");
 389                         goto err;
 390                 }
 391         }
 392
 393         if (tid && containerset) {
 394                 fprintf(stderr, "Select either -t or -C, not both\n");
 395                 goto err;
 396         }
 397
 398         /*
 399          * If we forked a child, wait for it to exit. Cannot use waitpid()
 400          * as all the delicious data would be reaped as part of the wait
 401          */
 402         if (tid && forking) {
 403                 int sig_received;
 404                 sigwait(&sigset, &sig_received);
 405         }
 406
 407         if (tid) {
 408                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 409                               cmd_type, &tid, sizeof(__u32));
 410                 PRINTF("Sent pid/tgid, retval %d\n", rc);
 411                 if (rc < 0) {
 412                         fprintf(stderr, "error sending tid/tgid cmd\n");
 413                         goto done;
 414                 }
 415         }
 416
 417         if (containerset) {
 418                 cfd = open(containerpath, O_RDONLY);
 419                 if (cfd < 0) {
 420                         perror("error opening container file");
 421                         goto err;
 422                 }
 423                 rc = send_cmd(nl_sd, id, mypid, CGROUPSTATS_CMD_GET,
 424                               CGROUPSTATS_CMD_ATTR_FD, &cfd, sizeof(__u32));
 425                 if (rc < 0) {
 426                         perror("error sending cgroupstats command");
 427                         goto err;
 428                 }
 429         }
 430         if (!maskset && !tid && !containerset) {
 431                 usage();
 432                 goto err;
 433         }
 434
 435         do {
 436                 int i;
 437
 438                 rep_len = recv(nl_sd, &msg, sizeof(msg), 0);
 439                 PRINTF("received %d bytes\n", rep_len);
 440
 441                 if (rep_len < 0) {
 442                         fprintf(stderr, "nonfatal reply error: errno %d\n",
 443                                 errno);
 444                         continue;
 445                 }
 446                 if (msg.n.nlmsg_type == NLMSG_ERROR ||
 447                     !NLMSG_OK((&msg.n), rep_len)) {
 448                         struct nlmsgerr *err = NLMSG_DATA(&msg);
 449                         fprintf(stderr, "fatal reply error,  errno %d\n",
 450                                 err->error);
 451                         goto done;
 452                 }
 453
 454                 PRINTF("nlmsghdr size=%zu, nlmsg_len=%d, rep_len=%d\n",
 455                        sizeof(struct nlmsghdr), msg.n.nlmsg_len, rep_len);
 456
 457
 458                 rep_len = GENLMSG_PAYLOAD(&msg.n);
 459
 460                 na = (struct nlattr *) GENLMSG_DATA(&msg);
 461                 len = 0;
 462                 i = 0;
 463                 while (len < rep_len) {
 464                         len += NLA_ALIGN(na->nla_len);
 465                         switch (na->nla_type) {
 466                         case TASKSTATS_TYPE_AGGR_TGID:
 467                                 /* Fall through */
 468                         case TASKSTATS_TYPE_AGGR_PID:
 469                                 aggr_len = NLA_PAYLOAD(na->nla_len);
 470                                 len2 = 0;
 471                                 /* For nested attributes, na follows */
 472                                 na = (struct nlattr *) NLA_DATA(na);
 473                                 done = 0;
 474                                 while (len2 < aggr_len) {
 475                                         switch (na->nla_type) {
 476                                         case TASKSTATS_TYPE_PID:
 477                                                 rtid = *(int *) NLA_DATA(na);
 478                                                 if (print_delays)
 479                                                         printf("PID\t%d\n", rtid);
 480                                                 break;
 481                                         case TASKSTATS_TYPE_TGID:
 482                                                 rtid = *(int *) NLA_DATA(na);
 483                                                 if (print_delays)
 484                                                         printf("TGID\t%d\n", rtid);
 485                                                 break;
 486                                         case TASKSTATS_TYPE_STATS:
 487                                                 count++;
 488                                                 if (print_delays)
 489                                                         print_delayacct((struct taskstats *) NLA_DATA(na));
 490                                                 if (print_io_accounting)
 491                                                         print_ioacct((struct taskstats *) NLA_DATA(na));
 492                                                 if (print_task_context_switch_counts)
 493                                                         task_context_switch_counts((struct taskstats *) NLA_DATA(na));
 494                                                 if (fd) {
 495                                                         if (write(fd, NLA_DATA(na), na->nla_len) < 0) {
 496                                                                 err(1,"write error\n");
 497                                                         }
 498                                                 }
 499                                                 if (!loop)
 500                                                         goto done;
 501                                                 break;
 502                                         default:
 503                                                 fprintf(stderr, "Unknown nested"
 504                                                         " nla_type %d\n",
 505                                                         na->nla_type);
 506                                                 break;
 507                                         }
 508                                         len2 += NLA_ALIGN(na->nla_len);
 509                                         na = (struct nlattr *) ((char *) na + len2);
 510                                 }
 511                                 break;
 512
 513                         case CGROUPSTATS_TYPE_CGROUP_STATS:
 514                                 print_cgroupstats(NLA_DATA(na));
 515                                 break;
 516                         default:
 517                                 fprintf(stderr, "Unknown nla_type %d\n",
 518                                         na->nla_type);
 519                                 break;
 520                         }
 521                         na = (struct nlattr *) (GENLMSG_DATA(&msg) + len);
 522                 }
 523         } while (loop);
 524 done:
 525         if (maskset) {
 526                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 527                               TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK,
 528                               &cpumask, strlen(cpumask) + 1);
 529                 printf("Sent deregister mask, retval %d\n", rc);
 530                 if (rc < 0)
 531                         err(rc, "error sending deregister cpumask\n");
 532         }
 533 err:
 534         close(nl_sd);
 535         if (fd)
 536                 close(fd);
 537         if (cfd)
 538                 close(cfd);
 539         return 0;
 540 }