blob: 64e49f5e4b552d3d4e67dea8ca7eed5dcf6e589e [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);
Jens Axboe4dced402007-07-23 14:16:55 +020078 if (fd < 0)
Jens Axboefb7b71a2007-05-15 08:44:04 +020079 return 0;
Jens Axboefb7b71a2007-05-15 08:44:04 +020080
81 ret = read(fd, &t, sizeof(t));
82 close(fd);
83
84 if (ret < 0) {
85 perror("read blktrace");
86 return 0;
87 } else if (ret != sizeof(t)) {
88 log_err("fio: short read on blktrace file\n");
89 return 0;
90 }
91
92 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
93 return 1;
94
95 return 0;
96}
97
Jens Axboe5e6c2062007-05-16 14:40:29 +020098static int lookup_device(char *path, unsigned int maj, unsigned int min)
99{
100 struct dirent *dir;
101 struct stat st;
102 int found = 0;
103 DIR *D;
104
105 D = opendir(path);
106 if (!D)
107 return 0;
108
109 while ((dir = readdir(D)) != NULL) {
110 char full_path[256];
111
112 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
113 continue;
114
115 sprintf(full_path, "%s/%s", path, dir->d_name);
116 if (lstat(full_path, &st) == -1) {
117 perror("lstat");
118 break;
119 }
120
121 if (S_ISDIR(st.st_mode)) {
122 found = lookup_device(full_path, maj, min);
123 if (found) {
124 strcpy(path, full_path);
125 break;
126 }
127 }
128
129 if (!S_ISBLK(st.st_mode))
130 continue;
131
132 if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
133 strcpy(path, full_path);
134 found = 1;
135 break;
136 }
137 }
138
139 closedir(D);
140 return found;
141}
142
Jens Axboec69aa912007-05-22 15:51:50 +0200143#define FMINORBITS 20
144#define FMINORMASK ((1U << FMINORBITS) - 1)
145#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
146#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
Jens Axboeeeb9c2a2007-05-16 18:28:47 +0200147
Jens Axboe5e6c2062007-05-16 14:40:29 +0200148static void trace_add_file(struct thread_data *td, __u32 device)
149{
150 static unsigned int last_maj, last_min;
Jens Axboec69aa912007-05-22 15:51:50 +0200151 unsigned int maj = FMAJOR(device);
152 unsigned int min = FMINOR(device);
Jens Axboe5e6c2062007-05-16 14:40:29 +0200153 struct fio_file *f;
154 char dev[256];
155 unsigned int i;
156
157 if (last_maj == maj && last_min == min)
158 return;
159
160 last_maj = maj;
161 last_min = min;
162
163 /*
164 * check for this file in our list
165 */
166 for_each_file(td, f, i)
167 if (f->major == maj && f->minor == min)
168 return;
169
170 strcpy(dev, "/dev");
171 if (lookup_device(dev, maj, min))
172 add_file(td, dev);
173}
174
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200175/*
176 * Store blk_io_trace data in an ipo for later retrieval.
177 */
Jens Axboefdefd982007-05-15 10:12:26 +0200178static void store_ipo(struct thread_data *td, unsigned long long offset,
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200179 unsigned int bytes, int rw, unsigned long long ttime)
Jens Axboefdefd982007-05-15 10:12:26 +0200180{
181 struct io_piece *ipo = malloc(sizeof(*ipo));
182
183 memset(ipo, 0, sizeof(*ipo));
184 INIT_LIST_HEAD(&ipo->list);
Jens Axboea2eea812007-05-15 13:10:41 +0200185 /*
186 * the 512 is wrong here, it should be the hardware sector size...
187 */
188 ipo->offset = offset * 512;
Jens Axboefdefd982007-05-15 10:12:26 +0200189 ipo->len = bytes;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200190 ipo->delay = ttime / 1000;
Jens Axboefdefd982007-05-15 10:12:26 +0200191 if (rw)
192 ipo->ddir = DDIR_WRITE;
193 else
194 ipo->ddir = DDIR_READ;
195
196 list_add_tail(&ipo->list, &td->io_log_list);
197}
198
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200199/*
200 * We only care for queue traces, most of the others are side effects
201 * due to internal workings of the block layer.
202 */
203static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
Jens Axboed84f8d42007-05-16 08:47:46 +0200204 unsigned long long ttime, unsigned long *ios,
205 unsigned int *bs)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200206{
Jens Axboefdefd982007-05-15 10:12:26 +0200207 int rw;
208
209 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
210 return;
Jens Axboea2eea812007-05-15 13:10:41 +0200211 if (t->action & BLK_TC_ACT(BLK_TC_PC))
212 return;
Jens Axboe3f3d3612007-05-16 12:29:06 +0200213 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
Jens Axboea2eea812007-05-15 13:10:41 +0200214 return;
Jens Axboefdefd982007-05-15 10:12:26 +0200215
Jens Axboe5e6c2062007-05-16 14:40:29 +0200216 trace_add_file(td, t->device);
217
Jens Axboee7a7d702007-05-15 10:13:04 +0200218 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
Jens Axboed84f8d42007-05-16 08:47:46 +0200219
220 if (t->bytes > bs[rw])
221 bs[rw] = t->bytes;
222
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200223 ios[rw]++;
Jens Axboe6df8ada2007-05-15 13:23:19 +0200224 td->o.size += t->bytes;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200225 store_ipo(td, t->sector, t->bytes, rw, ttime);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200226}
227
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200228/*
229 * Load a blktrace file by reading all the blk_io_trace entries, and storing
230 * them as io_pieces like the fio text version would do.
231 */
Jens Axboefb7b71a2007-05-15 08:44:04 +0200232int load_blktrace(struct thread_data *td, const char *filename)
233{
Jens Axboea61edde2007-05-15 14:29:58 +0200234 unsigned long long ttime, delay;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200235 struct blk_io_trace t;
Jens Axboe4241ea82007-09-12 08:18:36 +0200236 unsigned long ios[2], skipped_writes;
Jens Axboea61edde2007-05-15 14:29:58 +0200237 unsigned int cpu;
Jens Axboed84f8d42007-05-16 08:47:46 +0200238 unsigned int rw_bs[2];
Jens Axboee2887562007-05-16 09:25:09 +0200239 struct fifo *fifo;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200240 int fd;
241
242 fd = open(filename, O_RDONLY);
243 if (fd < 0) {
244 td_verror(td, errno, "open blktrace file");
245 return 1;
246 }
247
Jens Axboee2887562007-05-16 09:25:09 +0200248 fifo = fifo_alloc(TRACE_FIFO_SIZE);
249
Jens Axboe6df8ada2007-05-15 13:23:19 +0200250 td->o.size = 0;
251
Jens Axboea61edde2007-05-15 14:29:58 +0200252 cpu = 0;
Jens Axboed84f8d42007-05-16 08:47:46 +0200253 ttime = 0;
254 ios[0] = ios[1] = 0;
255 rw_bs[0] = rw_bs[1] = 0;
Jens Axboe4241ea82007-09-12 08:18:36 +0200256 skipped_writes = 0;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200257 do {
Jens Axboee2887562007-05-16 09:25:09 +0200258 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
Jens Axboefb7b71a2007-05-15 08:44:04 +0200259
Jens Axboee2887562007-05-16 09:25:09 +0200260 if (ret < 0)
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200261 goto err;
Jens Axboee2887562007-05-16 09:25:09 +0200262 else if (!ret)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200263 break;
Jens Axboee2887562007-05-16 09:25:09 +0200264 else if (ret < (int) sizeof(t)) {
265 log_err("fio: short fifo get\n");
266 break;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200267 }
268
269 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
Jens Axboef12b3232007-05-16 12:16:44 +0200270 log_err("fio: bad magic in blktrace data: %x\n", t.magic);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200271 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200272 }
273 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
274 log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200275 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200276 }
Jens Axboef12b3232007-05-16 12:16:44 +0200277 ret = discard_pdu(td, fifo, fd, &t);
278 if (ret < 0) {
Jens Axboefb7b71a2007-05-15 08:44:04 +0200279 td_verror(td, ret, "blktrace lseek");
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200280 goto err;
Jens Axboef12b3232007-05-16 12:16:44 +0200281 } else if (t.pdu_len != ret) {
282 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
283 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200284 }
Jens Axboef3f72c52007-08-23 09:03:26 +0200285 if (t.action & BLK_TC_ACT(BLK_TC_NOTIFY))
286 continue;
Jens Axboea61edde2007-05-15 14:29:58 +0200287 if (!ttime) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200288 ttime = t.time;
Jens Axboea61edde2007-05-15 14:29:58 +0200289 cpu = t.cpu;
290 }
291 delay = 0;
292 if (cpu == t.cpu)
293 delay = t.time - ttime;
Jens Axboe4241ea82007-09-12 08:18:36 +0200294 if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
295 skipped_writes++;
296 else
297 handle_trace(td, &t, delay, ios, rw_bs);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200298 ttime = t.time;
Jens Axboea61edde2007-05-15 14:29:58 +0200299 cpu = t.cpu;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200300 } while (1);
301
Jens Axboe38470f82007-05-16 09:26:23 +0200302 fifo_free(fifo);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200303 close(fd);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200304
Jens Axboe4241ea82007-09-12 08:18:36 +0200305 if (skipped_writes)
306 log_err("fio: <%s> skips replay of %lu writes due to read-only\n", td->o.name, skipped_writes);
307
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200308 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
309 log_err("fio: found no ios in blktrace data\n");
310 return 1;
Jens Axboed84f8d42007-05-16 08:47:46 +0200311 } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200312 td->o.td_ddir = TD_DDIR_READ;
Jens Axboed84f8d42007-05-16 08:47:46 +0200313 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
314 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200315 td->o.td_ddir = TD_DDIR_WRITE;
Jens Axboed84f8d42007-05-16 08:47:46 +0200316 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
317 } else {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200318 td->o.td_ddir = TD_DDIR_RW;
Jens Axboed84f8d42007-05-16 08:47:46 +0200319 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
320 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
321 }
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200322
323 /*
324 * We need to do direct/raw ios to the device, to avoid getting
325 * read-ahead in our way.
326 */
327 td->o.odirect = 1;
328
Jens Axboefb7b71a2007-05-15 08:44:04 +0200329 return 0;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200330err:
331 close(fd);
Jens Axboe38470f82007-05-16 09:26:23 +0200332 fifo_free(fifo);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200333 return 1;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200334}