blob: 5ed2619d2cf845a33fa9fa2f349b3f0485f254de [file] [log] [blame]
Jens Axboefb7b71a2007-05-15 08:44:04 +02001/*
2 * blktrace support code for fio
3 */
4#include <stdio.h>
5#include <stdlib.h>
Jens Axboe5e6c2062007-05-16 14:40:29 +02006#include <sys/stat.h>
7#include <dirent.h>
Jens Axboe8c1fdf02007-05-15 11:54:21 +02008
Jens Axboefb7b71a2007-05-15 08:44:04 +02009#include "list.h"
10#include "fio.h"
11#include "blktrace_api.h"
12
Jens Axboef12b3232007-05-16 12:16:44 +020013#define TRACE_FIFO_SIZE 65536
Jens Axboee2887562007-05-16 09:25:09 +020014
15/*
16 * fifo refill frontend, to avoid reading data in trace sized bites
17 */
18static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
19{
20 char buf[TRACE_FIFO_SIZE];
Jens Axboef12b3232007-05-16 12:16:44 +020021 unsigned int total;
Jens Axboee2887562007-05-16 09:25:09 +020022 int ret;
23
Jens Axboef12b3232007-05-16 12:16:44 +020024 total = sizeof(buf);
25 if (total > fifo_room(fifo))
26 total = fifo_room(fifo);
Jens Axboee2887562007-05-16 09:25:09 +020027
Jens Axboef12b3232007-05-16 12:16:44 +020028 ret = read(fd, buf, total);
29 if (ret < 0) {
30 td_verror(td, errno, "read blktrace file");
31 return -1;
Jens Axboee2887562007-05-16 09:25:09 +020032 }
33
Jens Axboef12b3232007-05-16 12:16:44 +020034 if (ret > 0)
35 ret = fifo_put(fifo, buf, ret);
36
37 return ret;
Jens Axboee2887562007-05-16 09:25:09 +020038}
39
40/*
41 * Retrieve 'len' bytes from the fifo, refilling if necessary.
42 */
43static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
44 void *buf, unsigned int len)
45{
Jens Axboef12b3232007-05-16 12:16:44 +020046 if (fifo_len(fifo) < len) {
47 int ret = refill_fifo(td, fifo, fd);
Jens Axboee2887562007-05-16 09:25:09 +020048
Jens Axboef12b3232007-05-16 12:16:44 +020049 if (ret < 0)
50 return ret;
51 }
Jens Axboee2887562007-05-16 09:25:09 +020052
53 return fifo_get(fifo, buf, len);
54}
55
Jens Axboe8c1fdf02007-05-15 11:54:21 +020056/*
57 * Just discard the pdu by seeking past it.
58 */
Jens Axboef12b3232007-05-16 12:16:44 +020059static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
60 struct blk_io_trace *t)
Jens Axboefb7b71a2007-05-15 08:44:04 +020061{
62 if (t->pdu_len == 0)
63 return 0;
64
Jens Axboef12b3232007-05-16 12:16:44 +020065 return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
Jens Axboefb7b71a2007-05-15 08:44:04 +020066}
67
Jens Axboe8c1fdf02007-05-15 11:54:21 +020068/*
69 * Check if this is a blktrace binary data file. We read a single trace
70 * into memory and check for the magic signature.
71 */
Jens Axboefb7b71a2007-05-15 08:44:04 +020072int is_blktrace(const char *filename)
73{
74 struct blk_io_trace t;
75 int fd, ret;
76
77 fd = open(filename, O_RDONLY);
78 if (fd < 0) {
79 perror("open blktrace");
80 return 0;
81 }
82
83 ret = read(fd, &t, sizeof(t));
84 close(fd);
85
86 if (ret < 0) {
87 perror("read blktrace");
88 return 0;
89 } else if (ret != sizeof(t)) {
90 log_err("fio: short read on blktrace file\n");
91 return 0;
92 }
93
94 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
95 return 1;
96
97 return 0;
98}
99
Jens Axboe5e6c2062007-05-16 14:40:29 +0200100static int lookup_device(char *path, unsigned int maj, unsigned int min)
101{
102 struct dirent *dir;
103 struct stat st;
104 int found = 0;
105 DIR *D;
106
107 D = opendir(path);
108 if (!D)
109 return 0;
110
111 while ((dir = readdir(D)) != NULL) {
112 char full_path[256];
113
114 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
115 continue;
116
117 sprintf(full_path, "%s/%s", path, dir->d_name);
118 if (lstat(full_path, &st) == -1) {
119 perror("lstat");
120 break;
121 }
122
123 if (S_ISDIR(st.st_mode)) {
124 found = lookup_device(full_path, maj, min);
125 if (found) {
126 strcpy(path, full_path);
127 break;
128 }
129 }
130
131 if (!S_ISBLK(st.st_mode))
132 continue;
133
134 if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
135 strcpy(path, full_path);
136 found = 1;
137 break;
138 }
139 }
140
141 closedir(D);
142 return found;
143}
144
145static void trace_add_file(struct thread_data *td, __u32 device)
146{
147 static unsigned int last_maj, last_min;
148 unsigned int maj = major(device);
149 unsigned int min = minor(device);
150 struct fio_file *f;
151 char dev[256];
152 unsigned int i;
153
154 if (last_maj == maj && last_min == min)
155 return;
156
157 last_maj = maj;
158 last_min = min;
159
160 /*
161 * check for this file in our list
162 */
163 for_each_file(td, f, i)
164 if (f->major == maj && f->minor == min)
165 return;
166
167 strcpy(dev, "/dev");
168 if (lookup_device(dev, maj, min))
169 add_file(td, dev);
170}
171
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200172/*
173 * Store blk_io_trace data in an ipo for later retrieval.
174 */
Jens Axboefdefd982007-05-15 10:12:26 +0200175static void store_ipo(struct thread_data *td, unsigned long long offset,
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200176 unsigned int bytes, int rw, unsigned long long ttime)
Jens Axboefdefd982007-05-15 10:12:26 +0200177{
178 struct io_piece *ipo = malloc(sizeof(*ipo));
179
180 memset(ipo, 0, sizeof(*ipo));
181 INIT_LIST_HEAD(&ipo->list);
Jens Axboea2eea812007-05-15 13:10:41 +0200182 /*
183 * the 512 is wrong here, it should be the hardware sector size...
184 */
185 ipo->offset = offset * 512;
Jens Axboefdefd982007-05-15 10:12:26 +0200186 ipo->len = bytes;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200187 ipo->delay = ttime / 1000;
Jens Axboefdefd982007-05-15 10:12:26 +0200188 if (rw)
189 ipo->ddir = DDIR_WRITE;
190 else
191 ipo->ddir = DDIR_READ;
192
193 list_add_tail(&ipo->list, &td->io_log_list);
194}
195
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200196/*
197 * We only care for queue traces, most of the others are side effects
198 * due to internal workings of the block layer.
199 */
200static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
Jens Axboed84f8d42007-05-16 08:47:46 +0200201 unsigned long long ttime, unsigned long *ios,
202 unsigned int *bs)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200203{
Jens Axboefdefd982007-05-15 10:12:26 +0200204 int rw;
205
206 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
207 return;
Jens Axboea2eea812007-05-15 13:10:41 +0200208 if (t->action & BLK_TC_ACT(BLK_TC_PC))
209 return;
Jens Axboe3f3d3612007-05-16 12:29:06 +0200210 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
Jens Axboea2eea812007-05-15 13:10:41 +0200211 return;
Jens Axboefdefd982007-05-15 10:12:26 +0200212
Jens Axboe5e6c2062007-05-16 14:40:29 +0200213 trace_add_file(td, t->device);
214
Jens Axboee7a7d702007-05-15 10:13:04 +0200215 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
Jens Axboed84f8d42007-05-16 08:47:46 +0200216
217 if (t->bytes > bs[rw])
218 bs[rw] = t->bytes;
219
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200220 ios[rw]++;
Jens Axboe6df8ada2007-05-15 13:23:19 +0200221 td->o.size += t->bytes;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200222 store_ipo(td, t->sector, t->bytes, rw, ttime);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200223}
224
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200225/*
226 * Load a blktrace file by reading all the blk_io_trace entries, and storing
227 * them as io_pieces like the fio text version would do.
228 */
Jens Axboefb7b71a2007-05-15 08:44:04 +0200229int load_blktrace(struct thread_data *td, const char *filename)
230{
Jens Axboea61edde2007-05-15 14:29:58 +0200231 unsigned long long ttime, delay;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200232 struct blk_io_trace t;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200233 unsigned long ios[2];
Jens Axboea61edde2007-05-15 14:29:58 +0200234 unsigned int cpu;
Jens Axboed84f8d42007-05-16 08:47:46 +0200235 unsigned int rw_bs[2];
Jens Axboee2887562007-05-16 09:25:09 +0200236 struct fifo *fifo;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200237 int fd;
238
239 fd = open(filename, O_RDONLY);
240 if (fd < 0) {
241 td_verror(td, errno, "open blktrace file");
242 return 1;
243 }
244
Jens Axboee2887562007-05-16 09:25:09 +0200245 fifo = fifo_alloc(TRACE_FIFO_SIZE);
246
Jens Axboe6df8ada2007-05-15 13:23:19 +0200247 td->o.size = 0;
248
Jens Axboea61edde2007-05-15 14:29:58 +0200249 cpu = 0;
Jens Axboed84f8d42007-05-16 08:47:46 +0200250 ttime = 0;
251 ios[0] = ios[1] = 0;
252 rw_bs[0] = rw_bs[1] = 0;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200253 do {
Jens Axboee2887562007-05-16 09:25:09 +0200254 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
Jens Axboefb7b71a2007-05-15 08:44:04 +0200255
Jens Axboee2887562007-05-16 09:25:09 +0200256 if (ret < 0)
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200257 goto err;
Jens Axboee2887562007-05-16 09:25:09 +0200258 else if (!ret)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200259 break;
Jens Axboee2887562007-05-16 09:25:09 +0200260 else if (ret < (int) sizeof(t)) {
261 log_err("fio: short fifo get\n");
262 break;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200263 }
264
265 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
Jens Axboef12b3232007-05-16 12:16:44 +0200266 log_err("fio: bad magic in blktrace data: %x\n", t.magic);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200267 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200268 }
269 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
270 log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200271 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200272 }
Jens Axboef12b3232007-05-16 12:16:44 +0200273 ret = discard_pdu(td, fifo, fd, &t);
274 if (ret < 0) {
Jens Axboefb7b71a2007-05-15 08:44:04 +0200275 td_verror(td, ret, "blktrace lseek");
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200276 goto err;
Jens Axboef12b3232007-05-16 12:16:44 +0200277 } else if (t.pdu_len != ret) {
278 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
279 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200280 }
Jens Axboea61edde2007-05-15 14:29:58 +0200281 if (!ttime) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200282 ttime = t.time;
Jens Axboea61edde2007-05-15 14:29:58 +0200283 cpu = t.cpu;
284 }
285 delay = 0;
286 if (cpu == t.cpu)
287 delay = t.time - ttime;
Jens Axboed84f8d42007-05-16 08:47:46 +0200288 handle_trace(td, &t, delay, ios, rw_bs);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200289 ttime = t.time;
Jens Axboea61edde2007-05-15 14:29:58 +0200290 cpu = t.cpu;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200291 } while (1);
292
Jens Axboe38470f82007-05-16 09:26:23 +0200293 fifo_free(fifo);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200294 close(fd);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200295
296 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
297 log_err("fio: found no ios in blktrace data\n");
298 return 1;
Jens Axboed84f8d42007-05-16 08:47:46 +0200299 } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200300 td->o.td_ddir = TD_DDIR_READ;
Jens Axboed84f8d42007-05-16 08:47:46 +0200301 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
302 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200303 td->o.td_ddir = TD_DDIR_WRITE;
Jens Axboed84f8d42007-05-16 08:47:46 +0200304 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
305 } else {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200306 td->o.td_ddir = TD_DDIR_RW;
Jens Axboed84f8d42007-05-16 08:47:46 +0200307 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
308 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
309 }
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200310
311 /*
312 * We need to do direct/raw ios to the device, to avoid getting
313 * read-ahead in our way.
314 */
315 td->o.odirect = 1;
316
Jens Axboefb7b71a2007-05-15 08:44:04 +0200317 return 0;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200318err:
319 close(fd);
Jens Axboe38470f82007-05-16 09:26:23 +0200320 fifo_free(fifo);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200321 return 1;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200322}