taskstats: add context-switch counters
[linux-2.6.git] / 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/types.h>
25 #include <signal.h>
26
27 #include <linux/genetlink.h>
28 #include <linux/taskstats.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 }
83
84 /*
85  * Create a raw netlink socket and bind
86  */
87 static int create_nl_socket(int protocol)
88 {
89         int fd;
90         struct sockaddr_nl local;
91
92         fd = socket(AF_NETLINK, SOCK_RAW, protocol);
93         if (fd < 0)
94                 return -1;
95
96         if (rcvbufsz)
97                 if (setsockopt(fd, SOL_SOCKET, SO_RCVBUF,
98                                 &rcvbufsz, sizeof(rcvbufsz)) < 0) {
99                         fprintf(stderr, "Unable to set socket rcv buf size "
100                                         "to %d\n",
101                                 rcvbufsz);
102                         return -1;
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
118 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  */
162 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, 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
179         rep_len = recv(sd, &ans, sizeof(ans), 0);
180         if (ans.n.nlmsg_type == NLMSG_ERROR ||
181             (rep_len < 0) || !NLMSG_OK((&ans.n), rep_len))
182                 return 0;
183
184         na = (struct nlattr *) GENLMSG_DATA(&ans);
185         na = (struct nlattr *) ((char *) na + NLA_ALIGN(na->nla_len));
186         if (na->nla_type == CTRL_ATTR_FAMILY_ID) {
187                 id = *(__u16 *) NLA_DATA(na);
188         }
189         return id;
190 }
191
192 void print_delayacct(struct taskstats *t)
193 {
194         printf("\n\nCPU   %15s%15s%15s%15s\n"
195                "      %15llu%15llu%15llu%15llu\n"
196                "IO    %15s%15s\n"
197                "      %15llu%15llu\n"
198                "MEM   %15s%15s\n"
199                "      %15llu%15llu\n"
200                "count", "real total", "virtual total", "delay total",
201                t->cpu_count, t->cpu_run_real_total, t->cpu_run_virtual_total,
202                t->cpu_delay_total,
203                "count", "delay total",
204                t->blkio_count, t->blkio_delay_total,
205                "count", "delay total", t->swapin_count, t->swapin_delay_total);
206 }
207
208 void task_context_switch_counts(struct taskstats *t)
209 {
210         printf("\n\nTask   %15s%15s\n"
211                "       %15lu%15lu\n",
212                "voluntary", "nonvoluntary",
213                t->nvcsw, t->nivcsw);
214 }
215
216 void print_ioacct(struct taskstats *t)
217 {
218         printf("%s: read=%llu, write=%llu, cancelled_write=%llu\n",
219                 t->ac_comm,
220                 (unsigned long long)t->read_bytes,
221                 (unsigned long long)t->write_bytes,
222                 (unsigned long long)t->cancelled_write_bytes);
223 }
224
225 int main(int argc, char *argv[])
226 {
227         int c, rc, rep_len, aggr_len, len2, cmd_type;
228         __u16 id;
229         __u32 mypid;
230
231         struct nlattr *na;
232         int nl_sd = -1;
233         int len = 0;
234         pid_t tid = 0;
235         pid_t rtid = 0;
236
237         int fd = 0;
238         int count = 0;
239         int write_file = 0;
240         int maskset = 0;
241         char *logfile = NULL;
242         int loop = 0;
243
244         struct msgtemplate msg;
245
246         while (1) {
247                 c = getopt(argc, argv, "qdiw:r:m:t:p:vl");
248                 if (c < 0)
249                         break;
250
251                 switch (c) {
252                 case 'd':
253                         printf("print delayacct stats ON\n");
254                         print_delays = 1;
255                         break;
256                 case 'i':
257                         printf("printing IO accounting\n");
258                         print_io_accounting = 1;
259                         break;
260                 case 'q':
261                         printf("printing task/process context switch rates\n");
262                         print_task_context_switch_counts = 1;
263                         break;
264                 case 'w':
265                         logfile = strdup(optarg);
266                         printf("write to file %s\n", logfile);
267                         write_file = 1;
268                         break;
269                 case 'r':
270                         rcvbufsz = atoi(optarg);
271                         printf("receive buf size %d\n", rcvbufsz);
272                         if (rcvbufsz < 0)
273                                 err(1, "Invalid rcv buf size\n");
274                         break;
275                 case 'm':
276                         strncpy(cpumask, optarg, sizeof(cpumask));
277                         maskset = 1;
278                         printf("cpumask %s maskset %d\n", cpumask, maskset);
279                         break;
280                 case 't':
281                         tid = atoi(optarg);
282                         if (!tid)
283                                 err(1, "Invalid tgid\n");
284                         cmd_type = TASKSTATS_CMD_ATTR_TGID;
285                         break;
286                 case 'p':
287                         tid = atoi(optarg);
288                         if (!tid)
289                                 err(1, "Invalid pid\n");
290                         cmd_type = TASKSTATS_CMD_ATTR_PID;
291                         break;
292                 case 'v':
293                         printf("debug on\n");
294                         dbg = 1;
295                         break;
296                 case 'l':
297                         printf("listen forever\n");
298                         loop = 1;
299                         break;
300                 default:
301                         usage();
302                         exit(-1);
303                 }
304         }
305
306         if (write_file) {
307                 fd = open(logfile, O_WRONLY | O_CREAT | O_TRUNC,
308                           S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
309                 if (fd == -1) {
310                         perror("Cannot open output file\n");
311                         exit(1);
312                 }
313         }
314
315         if ((nl_sd = create_nl_socket(NETLINK_GENERIC)) < 0)
316                 err(1, "error creating Netlink socket\n");
317
318
319         mypid = getpid();
320         id = get_family_id(nl_sd);
321         if (!id) {
322                 fprintf(stderr, "Error getting family id, errno %d\n", errno);
323                 goto err;
324         }
325         PRINTF("family id %d\n", id);
326
327         if (maskset) {
328                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
329                               TASKSTATS_CMD_ATTR_REGISTER_CPUMASK,
330                               &cpumask, strlen(cpumask) + 1);
331                 PRINTF("Sent register cpumask, retval %d\n", rc);
332                 if (rc < 0) {
333                         fprintf(stderr, "error sending register cpumask\n");
334                         goto err;
335                 }
336         }
337
338         if (tid) {
339                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
340                               cmd_type, &tid, sizeof(__u32));
341                 PRINTF("Sent pid/tgid, retval %d\n", rc);
342                 if (rc < 0) {
343                         fprintf(stderr, "error sending tid/tgid cmd\n");
344                         goto done;
345                 }
346         }
347
348         do {
349                 int i;
350
351                 rep_len = recv(nl_sd, &msg, sizeof(msg), 0);
352                 PRINTF("received %d bytes\n", rep_len);
353
354                 if (rep_len < 0) {
355                         fprintf(stderr, "nonfatal reply error: errno %d\n",
356                                 errno);
357                         continue;
358                 }
359                 if (msg.n.nlmsg_type == NLMSG_ERROR ||
360                     !NLMSG_OK((&msg.n), rep_len)) {
361                         struct nlmsgerr *err = NLMSG_DATA(&msg);
362                         fprintf(stderr, "fatal reply error,  errno %d\n",
363                                 err->error);
364                         goto done;
365                 }
366
367                 PRINTF("nlmsghdr size=%d, nlmsg_len=%d, rep_len=%d\n",
368                        sizeof(struct nlmsghdr), msg.n.nlmsg_len, rep_len);
369
370
371                 rep_len = GENLMSG_PAYLOAD(&msg.n);
372
373                 na = (struct nlattr *) GENLMSG_DATA(&msg);
374                 len = 0;
375                 i = 0;
376                 while (len < rep_len) {
377                         len += NLA_ALIGN(na->nla_len);
378                         switch (na->nla_type) {
379                         case TASKSTATS_TYPE_AGGR_TGID:
380                                 /* Fall through */
381                         case TASKSTATS_TYPE_AGGR_PID:
382                                 aggr_len = NLA_PAYLOAD(na->nla_len);
383                                 len2 = 0;
384                                 /* For nested attributes, na follows */
385                                 na = (struct nlattr *) NLA_DATA(na);
386                                 done = 0;
387                                 while (len2 < aggr_len) {
388                                         switch (na->nla_type) {
389                                         case TASKSTATS_TYPE_PID:
390                                                 rtid = *(int *) NLA_DATA(na);
391                                                 if (print_delays)
392                                                         printf("PID\t%d\n", rtid);
393                                                 break;
394                                         case TASKSTATS_TYPE_TGID:
395                                                 rtid = *(int *) NLA_DATA(na);
396                                                 if (print_delays)
397                                                         printf("TGID\t%d\n", rtid);
398                                                 break;
399                                         case TASKSTATS_TYPE_STATS:
400                                                 count++;
401                                                 if (print_delays)
402                                                         print_delayacct((struct taskstats *) NLA_DATA(na));
403                                                 if (print_io_accounting)
404                                                         print_ioacct((struct taskstats *) NLA_DATA(na));
405                                                 if (print_task_context_switch_counts)
406                                                         task_context_switch_counts((struct taskstats *) NLA_DATA(na));
407                                                 if (fd) {
408                                                         if (write(fd, NLA_DATA(na), na->nla_len) < 0) {
409                                                                 err(1,"write error\n");
410                                                         }
411                                                 }
412                                                 if (!loop)
413                                                         goto done;
414                                                 break;
415                                         default:
416                                                 fprintf(stderr, "Unknown nested"
417                                                         " nla_type %d\n",
418                                                         na->nla_type);
419                                                 break;
420                                         }
421                                         len2 += NLA_ALIGN(na->nla_len);
422                                         na = (struct nlattr *) ((char *) na + len2);
423                                 }
424                                 break;
425
426                         default:
427                                 fprintf(stderr, "Unknown nla_type %d\n",
428                                         na->nla_type);
429                                 break;
430                         }
431                         na = (struct nlattr *) (GENLMSG_DATA(&msg) + len);
432                 }
433         } while (loop);
434 done:
435         if (maskset) {
436                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
437                               TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK,
438                               &cpumask, strlen(cpumask) + 1);
439                 printf("Sent deregister mask, retval %d\n", rc);
440                 if (rc < 0)
441                         err(rc, "error sending deregister cpumask\n");
442         }
443 err:
444         close(nl_sd);
445         if (fd)
446                 close(fd);
447         return 0;
448 }