[PATCH] getdelays: various fixes
[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 __u64 stime, utime;
52
53 #define PRINTF(fmt, arg...) {                   \
54             if (dbg) {                          \
55                 printf(fmt, ##arg);             \
56             }                                   \
57         }
58
59 /* Maximum size of response requested or message sent */
60 #define MAX_MSG_SIZE    1024
61 /* Maximum number of cpus expected to be specified in a cpumask */
62 #define MAX_CPUS        32
63 /* Maximum length of pathname to log file */
64 #define MAX_FILENAME    256
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 /*
75  * Create a raw netlink socket and bind
76  */
77 static int create_nl_socket(int protocol)
78 {
79         int fd;
80         struct sockaddr_nl local;
81
82         fd = socket(AF_NETLINK, SOCK_RAW, protocol);
83         if (fd < 0)
84                 return -1;
85
86         if (rcvbufsz)
87                 if (setsockopt(fd, SOL_SOCKET, SO_RCVBUF,
88                                 &rcvbufsz, sizeof(rcvbufsz)) < 0) {
89                         fprintf(stderr, "Unable to set socket rcv buf size "
90                                         "to %d\n",
91                                 rcvbufsz);
92                         return -1;
93                 }
94
95         memset(&local, 0, sizeof(local));
96         local.nl_family = AF_NETLINK;
97
98         if (bind(fd, (struct sockaddr *) &local, sizeof(local)) < 0)
99                 goto error;
100
101         return fd;
102 error:
103         close(fd);
104         return -1;
105 }
106
107
108 int send_cmd(int sd, __u16 nlmsg_type, __u32 nlmsg_pid,
109              __u8 genl_cmd, __u16 nla_type,
110              void *nla_data, int nla_len)
111 {
112         struct nlattr *na;
113         struct sockaddr_nl nladdr;
114         int r, buflen;
115         char *buf;
116
117         struct msgtemplate msg;
118
119         msg.n.nlmsg_len = NLMSG_LENGTH(GENL_HDRLEN);
120         msg.n.nlmsg_type = nlmsg_type;
121         msg.n.nlmsg_flags = NLM_F_REQUEST;
122         msg.n.nlmsg_seq = 0;
123         msg.n.nlmsg_pid = nlmsg_pid;
124         msg.g.cmd = genl_cmd;
125         msg.g.version = 0x1;
126         na = (struct nlattr *) GENLMSG_DATA(&msg);
127         na->nla_type = nla_type;
128         na->nla_len = nla_len + 1 + NLA_HDRLEN;
129         memcpy(NLA_DATA(na), nla_data, nla_len);
130         msg.n.nlmsg_len += NLMSG_ALIGN(na->nla_len);
131
132         buf = (char *) &msg;
133         buflen = msg.n.nlmsg_len ;
134         memset(&nladdr, 0, sizeof(nladdr));
135         nladdr.nl_family = AF_NETLINK;
136         while ((r = sendto(sd, buf, buflen, 0, (struct sockaddr *) &nladdr,
137                            sizeof(nladdr))) < buflen) {
138                 if (r > 0) {
139                         buf += r;
140                         buflen -= r;
141                 } else if (errno != EAGAIN)
142                         return -1;
143         }
144         return 0;
145 }
146
147
148 /*
149  * Probe the controller in genetlink to find the family id
150  * for the TASKSTATS family
151  */
152 int get_family_id(int sd)
153 {
154         struct {
155                 struct nlmsghdr n;
156                 struct genlmsghdr g;
157                 char buf[256];
158         } ans;
159
160         int id, rc;
161         struct nlattr *na;
162         int rep_len;
163
164         strcpy(name, TASKSTATS_GENL_NAME);
165         rc = send_cmd(sd, GENL_ID_CTRL, getpid(), CTRL_CMD_GETFAMILY,
166                         CTRL_ATTR_FAMILY_NAME, (void *)name,
167                         strlen(TASKSTATS_GENL_NAME)+1);
168
169         rep_len = recv(sd, &ans, sizeof(ans), 0);
170         if (ans.n.nlmsg_type == NLMSG_ERROR ||
171             (rep_len < 0) || !NLMSG_OK((&ans.n), rep_len))
172                 return 0;
173
174         na = (struct nlattr *) GENLMSG_DATA(&ans);
175         na = (struct nlattr *) ((char *) na + NLA_ALIGN(na->nla_len));
176         if (na->nla_type == CTRL_ATTR_FAMILY_ID) {
177                 id = *(__u16 *) NLA_DATA(na);
178         }
179         return id;
180 }
181
182 void print_delayacct(struct taskstats *t)
183 {
184         printf("\n\nCPU   %15s%15s%15s%15s\n"
185                "      %15llu%15llu%15llu%15llu\n"
186                "IO    %15s%15s\n"
187                "      %15llu%15llu\n"
188                "MEM   %15s%15s\n"
189                "      %15llu%15llu\n\n",
190                "count", "real total", "virtual total", "delay total",
191                t->cpu_count, t->cpu_run_real_total, t->cpu_run_virtual_total,
192                t->cpu_delay_total,
193                "count", "delay total",
194                t->blkio_count, t->blkio_delay_total,
195                "count", "delay total", t->swapin_count, t->swapin_delay_total);
196 }
197
198 int main(int argc, char *argv[])
199 {
200         int c, rc, rep_len, aggr_len, len2, cmd_type;
201         __u16 id;
202         __u32 mypid;
203
204         struct nlattr *na;
205         int nl_sd = -1;
206         int len = 0;
207         pid_t tid = 0;
208         pid_t rtid = 0;
209
210         int fd = 0;
211         int count = 0;
212         int write_file = 0;
213         int maskset = 0;
214         char logfile[128];
215         int loop = 0;
216
217         struct msgtemplate msg;
218
219         while (1) {
220                 c = getopt(argc, argv, "dw:r:m:t:p:v:l");
221                 if (c < 0)
222                         break;
223
224                 switch (c) {
225                 case 'd':
226                         printf("print delayacct stats ON\n");
227                         print_delays = 1;
228                         break;
229                 case 'w':
230                         strncpy(logfile, optarg, MAX_FILENAME);
231                         printf("write to file %s\n", logfile);
232                         write_file = 1;
233                         break;
234                 case 'r':
235                         rcvbufsz = atoi(optarg);
236                         printf("receive buf size %d\n", rcvbufsz);
237                         if (rcvbufsz < 0)
238                                 err(1, "Invalid rcv buf size\n");
239                         break;
240                 case 'm':
241                         strncpy(cpumask, optarg, sizeof(cpumask));
242                         maskset = 1;
243                         printf("cpumask %s maskset %d\n", cpumask, maskset);
244                         break;
245                 case 't':
246                         tid = atoi(optarg);
247                         if (!tid)
248                                 err(1, "Invalid tgid\n");
249                         cmd_type = TASKSTATS_CMD_ATTR_TGID;
250                         print_delays = 1;
251                         break;
252                 case 'p':
253                         tid = atoi(optarg);
254                         if (!tid)
255                                 err(1, "Invalid pid\n");
256                         cmd_type = TASKSTATS_CMD_ATTR_PID;
257                         print_delays = 1;
258                         break;
259                 case 'v':
260                         printf("debug on\n");
261                         dbg = 1;
262                         break;
263                 case 'l':
264                         printf("listen forever\n");
265                         loop = 1;
266                         break;
267                 default:
268                         printf("Unknown option %d\n", c);
269                         exit(-1);
270                 }
271         }
272
273         if (write_file) {
274                 fd = open(logfile, O_WRONLY | O_CREAT | O_TRUNC,
275                           S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
276                 if (fd == -1) {
277                         perror("Cannot open output file\n");
278                         exit(1);
279                 }
280         }
281
282         if ((nl_sd = create_nl_socket(NETLINK_GENERIC)) < 0)
283                 err(1, "error creating Netlink socket\n");
284
285
286         mypid = getpid();
287         id = get_family_id(nl_sd);
288         if (!id) {
289                 fprintf(stderr, "Error getting family id, errno %d\n", errno);
290                 goto err;
291         }
292         PRINTF("family id %d\n", id);
293
294         if (maskset) {
295                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
296                               TASKSTATS_CMD_ATTR_REGISTER_CPUMASK,
297                               &cpumask, strlen(cpumask) + 1);
298                 PRINTF("Sent register cpumask, retval %d\n", rc);
299                 if (rc < 0) {
300                         fprintf(stderr, "error sending register cpumask\n");
301                         goto err;
302                 }
303         }
304
305         if (tid) {
306                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
307                               cmd_type, &tid, sizeof(__u32));
308                 PRINTF("Sent pid/tgid, retval %d\n", rc);
309                 if (rc < 0) {
310                         fprintf(stderr, "error sending tid/tgid cmd\n");
311                         goto done;
312                 }
313         }
314
315         do {
316                 int i;
317
318                 rep_len = recv(nl_sd, &msg, sizeof(msg), 0);
319                 PRINTF("received %d bytes\n", rep_len);
320
321                 if (rep_len < 0) {
322                         fprintf(stderr, "nonfatal reply error: errno %d\n",
323                                 errno);
324                         continue;
325                 }
326                 if (msg.n.nlmsg_type == NLMSG_ERROR ||
327                     !NLMSG_OK((&msg.n), rep_len)) {
328                         struct nlmsgerr *err = NLMSG_DATA(&msg);
329                         fprintf(stderr, "fatal reply error,  errno %d\n",
330                                 err->error);
331                         goto done;
332                 }
333
334                 PRINTF("nlmsghdr size=%d, nlmsg_len=%d, rep_len=%d\n",
335                        sizeof(struct nlmsghdr), msg.n.nlmsg_len, rep_len);
336
337
338                 rep_len = GENLMSG_PAYLOAD(&msg.n);
339
340                 na = (struct nlattr *) GENLMSG_DATA(&msg);
341                 len = 0;
342                 i = 0;
343                 while (len < rep_len) {
344                         len += NLA_ALIGN(na->nla_len);
345                         switch (na->nla_type) {
346                         case TASKSTATS_TYPE_AGGR_TGID:
347                                 /* Fall through */
348                         case TASKSTATS_TYPE_AGGR_PID:
349                                 aggr_len = NLA_PAYLOAD(na->nla_len);
350                                 len2 = 0;
351                                 /* For nested attributes, na follows */
352                                 na = (struct nlattr *) NLA_DATA(na);
353                                 done = 0;
354                                 while (len2 < aggr_len) {
355                                         switch (na->nla_type) {
356                                         case TASKSTATS_TYPE_PID:
357                                                 rtid = *(int *) NLA_DATA(na);
358                                                 if (print_delays)
359                                                         printf("PID\t%d\n", rtid);
360                                                 break;
361                                         case TASKSTATS_TYPE_TGID:
362                                                 rtid = *(int *) NLA_DATA(na);
363                                                 if (print_delays)
364                                                         printf("TGID\t%d\n", rtid);
365                                                 break;
366                                         case TASKSTATS_TYPE_STATS:
367                                                 count++;
368                                                 if (print_delays)
369                                                         print_delayacct((struct taskstats *) NLA_DATA(na));
370                                                 if (fd) {
371                                                         if (write(fd, NLA_DATA(na), na->nla_len) < 0) {
372                                                                 err(1,"write error\n");
373                                                         }
374                                                 }
375                                                 if (!loop)
376                                                         goto done;
377                                                 break;
378                                         default:
379                                                 fprintf(stderr, "Unknown nested"
380                                                         " nla_type %d\n",
381                                                         na->nla_type);
382                                                 break;
383                                         }
384                                         len2 += NLA_ALIGN(na->nla_len);
385                                         na = (struct nlattr *) ((char *) na + len2);
386                                 }
387                                 break;
388
389                         default:
390                                 fprintf(stderr, "Unknown nla_type %d\n",
391                                         na->nla_type);
392                                 break;
393                         }
394                         na = (struct nlattr *) (GENLMSG_DATA(&msg) + len);
395                 }
396         } while (loop);
397 done:
398         if (maskset) {
399                 rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
400                               TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK,
401                               &cpumask, strlen(cpumask) + 1);
402                 printf("Sent deregister mask, retval %d\n", rc);
403                 if (rc < 0)
404                         err(rc, "error sending deregister cpumask\n");
405         }
406 err:
407         close(nl_sd);
408         if (fd)
409                 close(fd);
410         return 0;
411 }