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 <signal.h>
  25
  26 #include <linux/genetlink.h>
  27 #include <linux/taskstats.h>
  28 #include <linux/cgroupstats.h>
  29
  30 /*
  31  * Generic macros for dealing with netlink sockets. Might be duplicated
  32  * elsewhere. It is recommended that commercial grade applications use
  33  * libnl or libnetlink and use the interfaces provided by the library
  34  */
  35 #define GENLMSG_DATA(glh)       ((void *)(NLMSG_DATA(glh) + GENL_HDRLEN))
  36 #define GENLMSG_PAYLOAD(glh)    (NLMSG_PAYLOAD(glh, 0) - GENL_HDRLEN)
  37 #define NLA_DATA(na)            ((void *)((char*)(na) + NLA_HDRLEN))
  38 #define NLA_PAYLOAD(len)        (len - NLA_HDRLEN)
  39
  40 #define err(code, fmt, arg...)                  \
  41         do {                                    \
  42                 fprintf(stderr, fmt, ##arg);    \
  43                 exit(code);                     \
  44         } while (0)
  45
  46 int done;
  47 int rcvbufsz;
  48 char name[100];
  49 int dbg;
  50 int print_delays;
  51 int print_io_accounting;
  52 int print_task_context_switch_counts;
  53 __u64 stime, utime;
  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 "
 101                                         "to %d\n",
 102                                 rcvbufsz);
 103                         return -1;
 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 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 + 1 + 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 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
 180         rep_len = recv(sd, &ans, sizeof(ans), 0);
 181         if (ans.n.nlmsg_type == NLMSG_ERROR ||
 182             (rep_len < 0) || !NLMSG_OK((&ans.n), rep_len))
 183                 return 0;
 184
 185         na = (struct nlattr *) GENLMSG_DATA(&ans);
 186         na = (struct nlattr *) ((char *) na + NLA_ALIGN(na->nla_len));
 187         if (na->nla_type == CTRL_ATTR_FAMILY_ID) {
 188                 id = *(__u16 *) NLA_DATA(na);
 189         }
 190         return id;
 191 }
 192
 193 void print_delayacct(struct taskstats *t)
 194 {
 195         printf("\n\nCPU   %15s%15s%15s%15s\n"
 196                "      %15llu%15llu%15llu%15llu\n"
 197                "IO    %15s%15s\n"
 198                "      %15llu%15llu\n"
 199                "SWAP  %15s%15s\n"
 200                "      %15llu%15llu\n"
 201                "RECLAIM  %12s%15s\n"
 202                "      %15llu%15llu\n",
 203                "count", "real total", "virtual total", "delay total",
 204                t->cpu_count, t->cpu_run_real_total, t->cpu_run_virtual_total,
 205                t->cpu_delay_total,
 206                "count", "delay total",
 207                t->blkio_count, t->blkio_delay_total,
 208                "count", "delay total", t->swapin_count, t->swapin_delay_total,
 209                "count", "delay total",
 210                t->freepages_count, t->freepages_delay_total);
 211 }
 212
 213 void task_context_switch_counts(struct taskstats *t)
 214 {
 215         printf("\n\nTask   %15s%15s\n"
 216                "       %15llu%15llu\n",
 217                "voluntary", "nonvoluntary",
 218                t->nvcsw, t->nivcsw);
 219 }
 220
 221 void print_cgroupstats(struct cgroupstats *c)
 222 {
 223         printf("sleeping %llu, blocked %llu, running %llu, stopped %llu, "
 224                 "uninterruptible %llu\n", c->nr_sleeping, c->nr_io_wait,
 225                 c->nr_running, c->nr_stopped, c->nr_uninterruptible);
 226 }
 227
 228
 229 void print_ioacct(struct taskstats *t)
 230 {
 231         printf("%s: read=%llu, write=%llu, cancelled_write=%llu\n",
 232                 t->ac_comm,
 233                 (unsigned long long)t->read_bytes,
 234                 (unsigned long long)t->write_bytes,
 235                 (unsigned long long)t->cancelled_write_bytes);
 236 }
 237
 238 int main(int argc, char *argv[])
 239 {
 240         int c, rc, rep_len, aggr_len, len2, cmd_type;
 241         __u16 id;
 242         __u32 mypid;
 243
 244         struct nlattr *na;
 245         int nl_sd = -1;
 246         int len = 0;
 247         pid_t tid = 0;
 248         pid_t rtid = 0;
 249
 250         int fd = 0;
 251         int count = 0;
 252         int write_file = 0;
 253         int maskset = 0;
 254         char *logfile = NULL;
 255         int loop = 0;
 256         int containerset = 0;
 257         char containerpath[1024];
 258         int cfd = 0;
 259
 260         struct msgtemplate msg;
 261
 262         while (1) {
 263                 c = getopt(argc, argv, "qdiw:r:m:t:p:vlC:");
 264                 if (c < 0)
 265                         break;
 266
 267                 switch (c) {
 268                 case 'd':
 269                         printf("print delayacct stats ON\n");
 270                         print_delays = 1;
 271                         break;
 272                 case 'i':
 273                         printf("printing IO accounting\n");
 274                         print_io_accounting = 1;
 275                         break;
 276                 case 'q':
 277                         printf("printing task/process context switch rates\n");
 278                         print_task_context_switch_counts = 1;
 279                         break;
 280                 case 'C':
 281                         containerset = 1;
 282                         strncpy(containerpath, optarg, strlen(optarg) + 1);
 283                         break;
 284                 case 'w':
 285                         logfile = strdup(optarg);
 286                         printf("write to file %s\n", logfile);
 287                         write_file = 1;
 288                         break;
 289                 case 'r':
 290                         rcvbufsz = atoi(optarg);
 291                         printf("receive buf size %d\n", rcvbufsz);
 292                         if (rcvbufsz < 0)
 293                                 err(1, "Invalid rcv buf size\n");
 294                         break;
 295                 case 'm':
 296                         strncpy(cpumask, optarg, sizeof(cpumask));
 297                         maskset = 1;
 298                         printf("cpumask %s maskset %d\n", cpumask, maskset);
 299                         break;
 300                 case 't':
 301                         tid = atoi(optarg);
 302                         if (!tid)
 303                                 err(1, "Invalid tgid\n");
 304                         cmd_type = TASKSTATS_CMD_ATTR_TGID;
 305                         break;
 306                 case 'p':
 307                         tid = atoi(optarg);
 308                         if (!tid)
 309                                 err(1, "Invalid pid\n");
 310                         cmd_type = TASKSTATS_CMD_ATTR_PID;
 311                         break;
 312                 case 'v':
 313                         printf("debug on\n");
 314                         dbg = 1;
 315                         break;
 316                 case 'l':
 317                         printf("listen forever\n");
 318                         loop = 1;
 319                         break;
 320                 default:
 321                         usage();
 322                         exit(-1);
 323                 }
 324         }
 325
 326         if (write_file) {
 327                 fd = open(logfile, O_WRONLY | O_CREAT | O_TRUNC,
 328                           S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
 329                 if (fd == -1) {
 330                         perror("Cannot open output file\n");
 331                         exit(1);
 332                 }
 333         }
 334
 335         if ((nl_sd = create_nl_socket(NETLINK_GENERIC)) < 0)
 336                 err(1, "error creating Netlink socket\n");
 337
 338
 339         mypid = getpid();
 340         id = get_family_id(nl_sd);
 341         if (!id) {
 342                 fprintf(stderr, "Error getting family id, errno %d\n", errno);
 343                 goto err;
 344         }
 345         PRINTF("family id %d\n", id);
 346
 347         if (maskset) {
 348                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 349                               TASKSTATS_CMD_ATTR_REGISTER_CPUMASK,
 350                               &cpumask, strlen(cpumask) + 1);
 351                 PRINTF("Sent register cpumask, retval %d\n", rc);
 352                 if (rc < 0) {
 353                         fprintf(stderr, "error sending register cpumask\n");
 354                         goto err;
 355                 }
 356         }
 357
 358         if (tid && containerset) {
 359                 fprintf(stderr, "Select either -t or -C, not both\n");
 360                 goto err;
 361         }
 362
 363         if (tid) {
 364                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 365                               cmd_type, &tid, sizeof(__u32));
 366                 PRINTF("Sent pid/tgid, retval %d\n", rc);
 367                 if (rc < 0) {
 368                         fprintf(stderr, "error sending tid/tgid cmd\n");
 369                         goto done;
 370                 }
 371         }
 372
 373         if (containerset) {
 374                 cfd = open(containerpath, O_RDONLY);
 375                 if (cfd < 0) {
 376                         perror("error opening container file");
 377                         goto err;
 378                 }
 379                 rc = send_cmd(nl_sd, id, mypid, CGROUPSTATS_CMD_GET,
 380                               CGROUPSTATS_CMD_ATTR_FD, &cfd, sizeof(__u32));
 381                 if (rc < 0) {
 382                         perror("error sending cgroupstats command");
 383                         goto err;
 384                 }
 385         }
 386
 387         do {
 388                 int i;
 389
 390                 rep_len = recv(nl_sd, &msg, sizeof(msg), 0);
 391                 PRINTF("received %d bytes\n", rep_len);
 392
 393                 if (rep_len < 0) {
 394                         fprintf(stderr, "nonfatal reply error: errno %d\n",
 395                                 errno);
 396                         continue;
 397                 }
 398                 if (msg.n.nlmsg_type == NLMSG_ERROR ||
 399                     !NLMSG_OK((&msg.n), rep_len)) {
 400                         struct nlmsgerr *err = NLMSG_DATA(&msg);
 401                         fprintf(stderr, "fatal reply error,  errno %d\n",
 402                                 err->error);
 403                         goto done;
 404                 }
 405
 406                 PRINTF("nlmsghdr size=%zu, nlmsg_len=%d, rep_len=%d\n",
 407                        sizeof(struct nlmsghdr), msg.n.nlmsg_len, rep_len);
 408
 409
 410                 rep_len = GENLMSG_PAYLOAD(&msg.n);
 411
 412                 na = (struct nlattr *) GENLMSG_DATA(&msg);
 413                 len = 0;
 414                 i = 0;
 415                 while (len < rep_len) {
 416                         len += NLA_ALIGN(na->nla_len);
 417                         switch (na->nla_type) {
 418                         case TASKSTATS_TYPE_AGGR_TGID:
 419                                 /* Fall through */
 420                         case TASKSTATS_TYPE_AGGR_PID:
 421                                 aggr_len = NLA_PAYLOAD(na->nla_len);
 422                                 len2 = 0;
 423                                 /* For nested attributes, na follows */
 424                                 na = (struct nlattr *) NLA_DATA(na);
 425                                 done = 0;
 426                                 while (len2 < aggr_len) {
 427                                         switch (na->nla_type) {
 428                                         case TASKSTATS_TYPE_PID:
 429                                                 rtid = *(int *) NLA_DATA(na);
 430                                                 if (print_delays)
 431                                                         printf("PID\t%d\n", rtid);
 432                                                 break;
 433                                         case TASKSTATS_TYPE_TGID:
 434                                                 rtid = *(int *) NLA_DATA(na);
 435                                                 if (print_delays)
 436                                                         printf("TGID\t%d\n", rtid);
 437                                                 break;
 438                                         case TASKSTATS_TYPE_STATS:
 439                                                 count++;
 440                                                 if (print_delays)
 441                                                         print_delayacct((struct taskstats *) NLA_DATA(na));
 442                                                 if (print_io_accounting)
 443                                                         print_ioacct((struct taskstats *) NLA_DATA(na));
 444                                                 if (print_task_context_switch_counts)
 445                                                         task_context_switch_counts((struct taskstats *) NLA_DATA(na));
 446                                                 if (fd) {
 447                                                         if (write(fd, NLA_DATA(na), na->nla_len) < 0) {
 448                                                                 err(1,"write error\n");
 449                                                         }
 450                                                 }
 451                                                 if (!loop)
 452                                                         goto done;
 453                                                 break;
 454                                         default:
 455                                                 fprintf(stderr, "Unknown nested"
 456                                                         " nla_type %d\n",
 457                                                         na->nla_type);
 458                                                 break;
 459                                         }
 460                                         len2 += NLA_ALIGN(na->nla_len);
 461                                         na = (struct nlattr *) ((char *) na + len2);
 462                                 }
 463                                 break;
 464
 465                         case CGROUPSTATS_TYPE_CGROUP_STATS:
 466                                 print_cgroupstats(NLA_DATA(na));
 467                                 break;
 468                         default:
 469                                 fprintf(stderr, "Unknown nla_type %d\n",
 470                                         na->nla_type);
 471                                 break;
 472                         }
 473                         na = (struct nlattr *) (GENLMSG_DATA(&msg) + len);
 474                 }
 475         } while (loop);
 476 done:
 477         if (maskset) {
 478                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
 479                               TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK,
 480                               &cpumask, strlen(cpumask) + 1);
 481                 printf("Sent deregister mask, retval %d\n", rc);
 482                 if (rc < 0)
 483                         err(rc, "error sending deregister cpumask\n");
 484         }
 485 err:
 486         close(nl_sd);
 487         if (fd)
 488                 close(fd);
 489         if (cfd)
 490                 close(cfd);
 491         return 0;
 492 }