Merge branch 'upstream-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mfashe...
[linux-2.6] / 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 __u64 stime, utime;
53
54 #define PRINTF(fmt, arg...) {                   \
55             if (dbg) {                          \
56                 printf(fmt, ##arg);             \
57             }                                   \
58         }
59
60 /* Maximum size of response requested or message sent */
61 #define MAX_MSG_SIZE    1024
62 /* Maximum number of cpus expected to be specified in a cpumask */
63 #define MAX_CPUS        32
64 /* Maximum length of pathname to log file */
65 #define MAX_FILENAME    256
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 /*
76  * Create a raw netlink socket and bind
77  */
78 static int create_nl_socket(int protocol)
79 {
80         int fd;
81         struct sockaddr_nl local;
82
83         fd = socket(AF_NETLINK, SOCK_RAW, protocol);
84         if (fd < 0)
85                 return -1;
86
87         if (rcvbufsz)
88                 if (setsockopt(fd, SOL_SOCKET, SO_RCVBUF,
89                                 &rcvbufsz, sizeof(rcvbufsz)) < 0) {
90                         fprintf(stderr, "Unable to set socket rcv buf size "
91                                         "to %d\n",
92                                 rcvbufsz);
93                         return -1;
94                 }
95
96         memset(&local, 0, sizeof(local));
97         local.nl_family = AF_NETLINK;
98
99         if (bind(fd, (struct sockaddr *) &local, sizeof(local)) < 0)
100                 goto error;
101
102         return fd;
103 error:
104         close(fd);
105         return -1;
106 }
107
108
109 int send_cmd(int sd, __u16 nlmsg_type, __u32 nlmsg_pid,
110              __u8 genl_cmd, __u16 nla_type,
111              void *nla_data, int nla_len)
112 {
113         struct nlattr *na;
114         struct sockaddr_nl nladdr;
115         int r, buflen;
116         char *buf;
117
118         struct msgtemplate msg;
119
120         msg.n.nlmsg_len = NLMSG_LENGTH(GENL_HDRLEN);
121         msg.n.nlmsg_type = nlmsg_type;
122         msg.n.nlmsg_flags = NLM_F_REQUEST;
123         msg.n.nlmsg_seq = 0;
124         msg.n.nlmsg_pid = nlmsg_pid;
125         msg.g.cmd = genl_cmd;
126         msg.g.version = 0x1;
127         na = (struct nlattr *) GENLMSG_DATA(&msg);
128         na->nla_type = nla_type;
129         na->nla_len = nla_len + 1 + NLA_HDRLEN;
130         memcpy(NLA_DATA(na), nla_data, nla_len);
131         msg.n.nlmsg_len += NLMSG_ALIGN(na->nla_len);
132
133         buf = (char *) &msg;
134         buflen = msg.n.nlmsg_len ;
135         memset(&nladdr, 0, sizeof(nladdr));
136         nladdr.nl_family = AF_NETLINK;
137         while ((r = sendto(sd, buf, buflen, 0, (struct sockaddr *) &nladdr,
138                            sizeof(nladdr))) < buflen) {
139                 if (r > 0) {
140                         buf += r;
141                         buflen -= r;
142                 } else if (errno != EAGAIN)
143                         return -1;
144         }
145         return 0;
146 }
147
148
149 /*
150  * Probe the controller in genetlink to find the family id
151  * for the TASKSTATS family
152  */
153 int get_family_id(int sd)
154 {
155         struct {
156                 struct nlmsghdr n;
157                 struct genlmsghdr g;
158                 char buf[256];
159         } ans;
160
161         int id, rc;
162         struct nlattr *na;
163         int rep_len;
164
165         strcpy(name, TASKSTATS_GENL_NAME);
166         rc = send_cmd(sd, GENL_ID_CTRL, getpid(), CTRL_CMD_GETFAMILY,
167                         CTRL_ATTR_FAMILY_NAME, (void *)name,
168                         strlen(TASKSTATS_GENL_NAME)+1);
169
170         rep_len = recv(sd, &ans, sizeof(ans), 0);
171         if (ans.n.nlmsg_type == NLMSG_ERROR ||
172             (rep_len < 0) || !NLMSG_OK((&ans.n), rep_len))
173                 return 0;
174
175         na = (struct nlattr *) GENLMSG_DATA(&ans);
176         na = (struct nlattr *) ((char *) na + NLA_ALIGN(na->nla_len));
177         if (na->nla_type == CTRL_ATTR_FAMILY_ID) {
178                 id = *(__u16 *) NLA_DATA(na);
179         }
180         return id;
181 }
182
183 void print_delayacct(struct taskstats *t)
184 {
185         printf("\n\nCPU   %15s%15s%15s%15s\n"
186                "      %15llu%15llu%15llu%15llu\n"
187                "IO    %15s%15s\n"
188                "      %15llu%15llu\n"
189                "MEM   %15s%15s\n"
190                "      %15llu%15llu\n\n",
191                "count", "real total", "virtual total", "delay total",
192                t->cpu_count, t->cpu_run_real_total, t->cpu_run_virtual_total,
193                t->cpu_delay_total,
194                "count", "delay total",
195                t->blkio_count, t->blkio_delay_total,
196                "count", "delay total", t->swapin_count, t->swapin_delay_total);
197 }
198
199 void print_ioacct(struct taskstats *t)
200 {
201         printf("%s: read=%llu, write=%llu, cancelled_write=%llu\n",
202                 t->ac_comm,
203                 (unsigned long long)t->read_bytes,
204                 (unsigned long long)t->write_bytes,
205                 (unsigned long long)t->cancelled_write_bytes);
206 }
207
208 int main(int argc, char *argv[])
209 {
210         int c, rc, rep_len, aggr_len, len2, cmd_type;
211         __u16 id;
212         __u32 mypid;
213
214         struct nlattr *na;
215         int nl_sd = -1;
216         int len = 0;
217         pid_t tid = 0;
218         pid_t rtid = 0;
219
220         int fd = 0;
221         int count = 0;
222         int write_file = 0;
223         int maskset = 0;
224         char logfile[128];
225         int loop = 0;
226
227         struct msgtemplate msg;
228
229         while (1) {
230                 c = getopt(argc, argv, "diw:r:m:t:p:v:l");
231                 if (c < 0)
232                         break;
233
234                 switch (c) {
235                 case 'd':
236                         printf("print delayacct stats ON\n");
237                         print_delays = 1;
238                         break;
239                 case 'i':
240                         printf("printing IO accounting\n");
241                         print_io_accounting = 1;
242                         break;
243                 case 'w':
244                         strncpy(logfile, optarg, MAX_FILENAME);
245                         printf("write to file %s\n", logfile);
246                         write_file = 1;
247                         break;
248                 case 'r':
249                         rcvbufsz = atoi(optarg);
250                         printf("receive buf size %d\n", rcvbufsz);
251                         if (rcvbufsz < 0)
252                                 err(1, "Invalid rcv buf size\n");
253                         break;
254                 case 'm':
255                         strncpy(cpumask, optarg, sizeof(cpumask));
256                         maskset = 1;
257                         printf("cpumask %s maskset %d\n", cpumask, maskset);
258                         break;
259                 case 't':
260                         tid = atoi(optarg);
261                         if (!tid)
262                                 err(1, "Invalid tgid\n");
263                         cmd_type = TASKSTATS_CMD_ATTR_TGID;
264                         break;
265                 case 'p':
266                         tid = atoi(optarg);
267                         if (!tid)
268                                 err(1, "Invalid pid\n");
269                         cmd_type = TASKSTATS_CMD_ATTR_PID;
270                         break;
271                 case 'v':
272                         printf("debug on\n");
273                         dbg = 1;
274                         break;
275                 case 'l':
276                         printf("listen forever\n");
277                         loop = 1;
278                         break;
279                 default:
280                         printf("Unknown option %d\n", c);
281                         exit(-1);
282                 }
283         }
284
285         if (write_file) {
286                 fd = open(logfile, O_WRONLY | O_CREAT | O_TRUNC,
287                           S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
288                 if (fd == -1) {
289                         perror("Cannot open output file\n");
290                         exit(1);
291                 }
292         }
293
294         if ((nl_sd = create_nl_socket(NETLINK_GENERIC)) < 0)
295                 err(1, "error creating Netlink socket\n");
296
297
298         mypid = getpid();
299         id = get_family_id(nl_sd);
300         if (!id) {
301                 fprintf(stderr, "Error getting family id, errno %d\n", errno);
302                 goto err;
303         }
304         PRINTF("family id %d\n", id);
305
306         if (maskset) {
307                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
308                               TASKSTATS_CMD_ATTR_REGISTER_CPUMASK,
309                               &cpumask, strlen(cpumask) + 1);
310                 PRINTF("Sent register cpumask, retval %d\n", rc);
311                 if (rc < 0) {
312                         fprintf(stderr, "error sending register cpumask\n");
313                         goto err;
314                 }
315         }
316
317         if (tid) {
318                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
319                               cmd_type, &tid, sizeof(__u32));
320                 PRINTF("Sent pid/tgid, retval %d\n", rc);
321                 if (rc < 0) {
322                         fprintf(stderr, "error sending tid/tgid cmd\n");
323                         goto done;
324                 }
325         }
326
327         do {
328                 int i;
329
330                 rep_len = recv(nl_sd, &msg, sizeof(msg), 0);
331                 PRINTF("received %d bytes\n", rep_len);
332
333                 if (rep_len < 0) {
334                         fprintf(stderr, "nonfatal reply error: errno %d\n",
335                                 errno);
336                         continue;
337                 }
338                 if (msg.n.nlmsg_type == NLMSG_ERROR ||
339                     !NLMSG_OK((&msg.n), rep_len)) {
340                         struct nlmsgerr *err = NLMSG_DATA(&msg);
341                         fprintf(stderr, "fatal reply error,  errno %d\n",
342                                 err->error);
343                         goto done;
344                 }
345
346                 PRINTF("nlmsghdr size=%d, nlmsg_len=%d, rep_len=%d\n",
347                        sizeof(struct nlmsghdr), msg.n.nlmsg_len, rep_len);
348
349
350                 rep_len = GENLMSG_PAYLOAD(&msg.n);
351
352                 na = (struct nlattr *) GENLMSG_DATA(&msg);
353                 len = 0;
354                 i = 0;
355                 while (len < rep_len) {
356                         len += NLA_ALIGN(na->nla_len);
357                         switch (na->nla_type) {
358                         case TASKSTATS_TYPE_AGGR_TGID:
359                                 /* Fall through */
360                         case TASKSTATS_TYPE_AGGR_PID:
361                                 aggr_len = NLA_PAYLOAD(na->nla_len);
362                                 len2 = 0;
363                                 /* For nested attributes, na follows */
364                                 na = (struct nlattr *) NLA_DATA(na);
365                                 done = 0;
366                                 while (len2 < aggr_len) {
367                                         switch (na->nla_type) {
368                                         case TASKSTATS_TYPE_PID:
369                                                 rtid = *(int *) NLA_DATA(na);
370                                                 if (print_delays)
371                                                         printf("PID\t%d\n", rtid);
372                                                 break;
373                                         case TASKSTATS_TYPE_TGID:
374                                                 rtid = *(int *) NLA_DATA(na);
375                                                 if (print_delays)
376                                                         printf("TGID\t%d\n", rtid);
377                                                 break;
378                                         case TASKSTATS_TYPE_STATS:
379                                                 count++;
380                                                 if (print_delays)
381                                                         print_delayacct((struct taskstats *) NLA_DATA(na));
382                                                 if (print_io_accounting)
383                                                         print_ioacct((struct taskstats *) NLA_DATA(na));
384                                                 if (fd) {
385                                                         if (write(fd, NLA_DATA(na), na->nla_len) < 0) {
386                                                                 err(1,"write error\n");
387                                                         }
388                                                 }
389                                                 if (!loop)
390                                                         goto done;
391                                                 break;
392                                         default:
393                                                 fprintf(stderr, "Unknown nested"
394                                                         " nla_type %d\n",
395                                                         na->nla_type);
396                                                 break;
397                                         }
398                                         len2 += NLA_ALIGN(na->nla_len);
399                                         na = (struct nlattr *) ((char *) na + len2);
400                                 }
401                                 break;
402
403                         default:
404                                 fprintf(stderr, "Unknown nla_type %d\n",
405                                         na->nla_type);
406                                 break;
407                         }
408                         na = (struct nlattr *) (GENLMSG_DATA(&msg) + len);
409                 }
410         } while (loop);
411 done:
412         if (maskset) {
413                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
414                               TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK,
415                               &cpumask, strlen(cpumask) + 1);
416                 printf("Sent deregister mask, retval %d\n", rc);
417                 if (rc < 0)
418                         err(rc, "error sending deregister cpumask\n");
419         }
420 err:
421         close(nl_sd);
422         if (fd)
423                 close(fd);
424         return 0;
425 }