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
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
usage(void)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 */
create_nl_socket(int protocol)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
send_cmd(int sd,__u16 nlmsg_type,__u32 nlmsg_pid,__u8 genl_cmd,__u16 nla_type,void * nla_data,int nla_len)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 + 1 + 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 */
get_family_id(int sd)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
print_delayacct(struct taskstats * t)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 "count", "real total", "virtual total",
207 "delay total", "delay average",
208 (unsigned long long)t->cpu_count,
209 (unsigned long long)t->cpu_run_real_total,
210 (unsigned long long)t->cpu_run_virtual_total,
211 (unsigned long long)t->cpu_delay_total,
212 average_ms((double)t->cpu_delay_total, t->cpu_count),
213 "count", "delay total", "delay average",
214 (unsigned long long)t->blkio_count,
215 (unsigned long long)t->blkio_delay_total,
216 average_ms(t->blkio_delay_total, t->blkio_count),
217 "count", "delay total", "delay average",
218 (unsigned long long)t->swapin_count,
219 (unsigned long long)t->swapin_delay_total,
220 average_ms(t->swapin_delay_total, t->swapin_count),
221 "count", "delay total", "delay average",
222 (unsigned long long)t->freepages_count,
223 (unsigned long long)t->freepages_delay_total,
224 average_ms(t->freepages_delay_total, t->freepages_count));
225 }
226
task_context_switch_counts(struct taskstats * t)227 static void task_context_switch_counts(struct taskstats *t)
228 {
229 printf("\n\nTask %15s%15s\n"
230 " %15llu%15llu\n",
231 "voluntary", "nonvoluntary",
232 (unsigned long long)t->nvcsw, (unsigned long long)t->nivcsw);
233 }
234
print_cgroupstats(struct cgroupstats * c)235 static void print_cgroupstats(struct cgroupstats *c)
236 {
237 printf("sleeping %llu, blocked %llu, running %llu, stopped %llu, "
238 "uninterruptible %llu\n", (unsigned long long)c->nr_sleeping,
239 (unsigned long long)c->nr_io_wait,
240 (unsigned long long)c->nr_running,
241 (unsigned long long)c->nr_stopped,
242 (unsigned long long)c->nr_uninterruptible);
243 }
244
245
print_ioacct(struct taskstats * t)246 static void print_ioacct(struct taskstats *t)
247 {
248 printf("%s: read=%llu, write=%llu, cancelled_write=%llu\n",
249 t->ac_comm,
250 (unsigned long long)t->read_bytes,
251 (unsigned long long)t->write_bytes,
252 (unsigned long long)t->cancelled_write_bytes);
253 }
254
main(int argc,char * argv[])255 int main(int argc, char *argv[])
256 {
257 int c, rc, rep_len, aggr_len, len2;
258 int cmd_type = TASKSTATS_CMD_ATTR_UNSPEC;
259 __u16 id;
260 __u32 mypid;
261
262 struct nlattr *na;
263 int nl_sd = -1;
264 int len = 0;
265 pid_t tid = 0;
266 pid_t rtid = 0;
267
268 int fd = 0;
269 int count = 0;
270 int write_file = 0;
271 int maskset = 0;
272 char *logfile = NULL;
273 int loop = 0;
274 int containerset = 0;
275 char containerpath[1024];
276 int cfd = 0;
277 int forking = 0;
278 sigset_t sigset;
279
280 struct msgtemplate msg;
281
282 while (!forking) {
283 c = getopt(argc, argv, "qdiw:r:m:t:p:vlC:c:");
284 if (c < 0)
285 break;
286
287 switch (c) {
288 case 'd':
289 printf("print delayacct stats ON\n");
290 print_delays = 1;
291 break;
292 case 'i':
293 printf("printing IO accounting\n");
294 print_io_accounting = 1;
295 break;
296 case 'q':
297 printf("printing task/process context switch rates\n");
298 print_task_context_switch_counts = 1;
299 break;
300 case 'C':
301 containerset = 1;
302 strncpy(containerpath, optarg, strlen(optarg) + 1);
303 break;
304 case 'w':
305 logfile = strdup(optarg);
306 printf("write to file %s\n", logfile);
307 write_file = 1;
308 break;
309 case 'r':
310 rcvbufsz = atoi(optarg);
311 printf("receive buf size %d\n", rcvbufsz);
312 if (rcvbufsz < 0)
313 err(1, "Invalid rcv buf size\n");
314 break;
315 case 'm':
316 strncpy(cpumask, optarg, sizeof(cpumask));
317 maskset = 1;
318 printf("cpumask %s maskset %d\n", cpumask, maskset);
319 break;
320 case 't':
321 tid = atoi(optarg);
322 if (!tid)
323 err(1, "Invalid tgid\n");
324 cmd_type = TASKSTATS_CMD_ATTR_TGID;
325 break;
326 case 'p':
327 tid = atoi(optarg);
328 if (!tid)
329 err(1, "Invalid pid\n");
330 cmd_type = TASKSTATS_CMD_ATTR_PID;
331 break;
332 case 'c':
333
334 /* Block SIGCHLD for sigwait() later */
335 if (sigemptyset(&sigset) == -1)
336 err(1, "Failed to empty sigset");
337 if (sigaddset(&sigset, SIGCHLD))
338 err(1, "Failed to set sigchld in sigset");
339 sigprocmask(SIG_BLOCK, &sigset, NULL);
340
341 /* fork/exec a child */
342 tid = fork();
343 if (tid < 0)
344 err(1, "Fork failed\n");
345 if (tid == 0)
346 if (execvp(argv[optind - 1],
347 &argv[optind - 1]) < 0)
348 exit(-1);
349
350 /* Set the command type and avoid further processing */
351 cmd_type = TASKSTATS_CMD_ATTR_PID;
352 forking = 1;
353 break;
354 case 'v':
355 printf("debug on\n");
356 dbg = 1;
357 break;
358 case 'l':
359 printf("listen forever\n");
360 loop = 1;
361 break;
362 default:
363 usage();
364 exit(-1);
365 }
366 }
367
368 if (write_file) {
369 fd = open(logfile, O_WRONLY | O_CREAT | O_TRUNC,
370 S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
371 if (fd == -1) {
372 perror("Cannot open output file\n");
373 exit(1);
374 }
375 }
376
377 if ((nl_sd = create_nl_socket(NETLINK_GENERIC)) < 0)
378 err(1, "error creating Netlink socket\n");
379
380
381 mypid = getpid();
382 id = get_family_id(nl_sd);
383 if (!id) {
384 fprintf(stderr, "Error getting family id, errno %d\n", errno);
385 goto err;
386 }
387 PRINTF("family id %d\n", id);
388
389 if (maskset) {
390 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
391 TASKSTATS_CMD_ATTR_REGISTER_CPUMASK,
392 &cpumask, strlen(cpumask) + 1);
393 PRINTF("Sent register cpumask, retval %d\n", rc);
394 if (rc < 0) {
395 fprintf(stderr, "error sending register cpumask\n");
396 goto err;
397 }
398 }
399
400 if (tid && containerset) {
401 fprintf(stderr, "Select either -t or -C, not both\n");
402 goto err;
403 }
404
405 /*
406 * If we forked a child, wait for it to exit. Cannot use waitpid()
407 * as all the delicious data would be reaped as part of the wait
408 */
409 if (tid && forking) {
410 int sig_received;
411 sigwait(&sigset, &sig_received);
412 }
413
414 if (tid) {
415 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
416 cmd_type, &tid, sizeof(__u32));
417 PRINTF("Sent pid/tgid, retval %d\n", rc);
418 if (rc < 0) {
419 fprintf(stderr, "error sending tid/tgid cmd\n");
420 goto done;
421 }
422 }
423
424 if (containerset) {
425 cfd = open(containerpath, O_RDONLY);
426 if (cfd < 0) {
427 perror("error opening container file");
428 goto err;
429 }
430 rc = send_cmd(nl_sd, id, mypid, CGROUPSTATS_CMD_GET,
431 CGROUPSTATS_CMD_ATTR_FD, &cfd, sizeof(__u32));
432 if (rc < 0) {
433 perror("error sending cgroupstats command");
434 goto err;
435 }
436 }
437 if (!maskset && !tid && !containerset) {
438 usage();
439 goto err;
440 }
441
442 do {
443 rep_len = recv(nl_sd, &msg, sizeof(msg), 0);
444 PRINTF("received %d bytes\n", rep_len);
445
446 if (rep_len < 0) {
447 fprintf(stderr, "nonfatal reply error: errno %d\n",
448 errno);
449 continue;
450 }
451 if (msg.n.nlmsg_type == NLMSG_ERROR ||
452 !NLMSG_OK((&msg.n), rep_len)) {
453 struct nlmsgerr *err = NLMSG_DATA(&msg);
454 fprintf(stderr, "fatal reply error, errno %d\n",
455 err->error);
456 goto done;
457 }
458
459 PRINTF("nlmsghdr size=%zu, nlmsg_len=%d, rep_len=%d\n",
460 sizeof(struct nlmsghdr), msg.n.nlmsg_len, rep_len);
461
462
463 rep_len = GENLMSG_PAYLOAD(&msg.n);
464
465 na = (struct nlattr *) GENLMSG_DATA(&msg);
466 len = 0;
467 while (len < rep_len) {
468 len += NLA_ALIGN(na->nla_len);
469 switch (na->nla_type) {
470 case TASKSTATS_TYPE_AGGR_TGID:
471 /* Fall through */
472 case TASKSTATS_TYPE_AGGR_PID:
473 aggr_len = NLA_PAYLOAD(na->nla_len);
474 len2 = 0;
475 /* For nested attributes, na follows */
476 na = (struct nlattr *) NLA_DATA(na);
477 done = 0;
478 while (len2 < aggr_len) {
479 switch (na->nla_type) {
480 case TASKSTATS_TYPE_PID:
481 rtid = *(int *) NLA_DATA(na);
482 if (print_delays)
483 printf("PID\t%d\n", rtid);
484 break;
485 case TASKSTATS_TYPE_TGID:
486 rtid = *(int *) NLA_DATA(na);
487 if (print_delays)
488 printf("TGID\t%d\n", rtid);
489 break;
490 case TASKSTATS_TYPE_STATS:
491 count++;
492 if (print_delays)
493 print_delayacct((struct taskstats *) NLA_DATA(na));
494 if (print_io_accounting)
495 print_ioacct((struct taskstats *) NLA_DATA(na));
496 if (print_task_context_switch_counts)
497 task_context_switch_counts((struct taskstats *) NLA_DATA(na));
498 if (fd) {
499 if (write(fd, NLA_DATA(na), na->nla_len) < 0) {
500 err(1,"write error\n");
501 }
502 }
503 if (!loop)
504 goto done;
505 break;
506 default:
507 fprintf(stderr, "Unknown nested"
508 " nla_type %d\n",
509 na->nla_type);
510 break;
511 }
512 len2 += NLA_ALIGN(na->nla_len);
513 na = (struct nlattr *) ((char *) na + len2);
514 }
515 break;
516
517 case CGROUPSTATS_TYPE_CGROUP_STATS:
518 print_cgroupstats(NLA_DATA(na));
519 break;
520 default:
521 fprintf(stderr, "Unknown nla_type %d\n",
522 na->nla_type);
523 case TASKSTATS_TYPE_NULL:
524 break;
525 }
526 na = (struct nlattr *) (GENLMSG_DATA(&msg) + len);
527 }
528 } while (loop);
529 done:
530 if (maskset) {
531 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
532 TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK,
533 &cpumask, strlen(cpumask) + 1);
534 printf("Sent deregister mask, retval %d\n", rc);
535 if (rc < 0)
536 err(rc, "error sending deregister cpumask\n");
537 }
538 err:
539 close(nl_sd);
540 if (fd)
541 close(fd);
542 if (cfd)
543 close(cfd);
544 return 0;
545 }
546