blob: 4469243efdb502f97314b81e61f614ab3f854aaf [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 Axboe01743ee2008-06-02 12:19:19 +02009#include "flist.h"
Jens Axboefb7b71a2007-05-15 08:44:04 +020010#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
Jens Axboebd6f78b2008-02-01 20:27:52 +010037 dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
Jens Axboef12b3232007-05-16 12:16:44 +020038 return ret;
Jens Axboee2887562007-05-16 09:25:09 +020039}
40
41/*
42 * Retrieve 'len' bytes from the fifo, refilling if necessary.
43 */
44static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
45 void *buf, unsigned int len)
46{
Jens Axboef12b3232007-05-16 12:16:44 +020047 if (fifo_len(fifo) < len) {
48 int ret = refill_fifo(td, fifo, fd);
Jens Axboee2887562007-05-16 09:25:09 +020049
Jens Axboef12b3232007-05-16 12:16:44 +020050 if (ret < 0)
51 return ret;
52 }
Jens Axboee2887562007-05-16 09:25:09 +020053
54 return fifo_get(fifo, buf, len);
55}
56
Jens Axboe8c1fdf02007-05-15 11:54:21 +020057/*
58 * Just discard the pdu by seeking past it.
59 */
Jens Axboef12b3232007-05-16 12:16:44 +020060static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
61 struct blk_io_trace *t)
Jens Axboefb7b71a2007-05-15 08:44:04 +020062{
63 if (t->pdu_len == 0)
64 return 0;
65
Jens Axboebd6f78b2008-02-01 20:27:52 +010066 dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
Jens Axboef12b3232007-05-16 12:16:44 +020067 return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
Jens Axboefb7b71a2007-05-15 08:44:04 +020068}
69
Jens Axboe8c1fdf02007-05-15 11:54:21 +020070/*
71 * Check if this is a blktrace binary data file. We read a single trace
72 * into memory and check for the magic signature.
73 */
Jens Axboefb7b71a2007-05-15 08:44:04 +020074int is_blktrace(const char *filename)
75{
76 struct blk_io_trace t;
77 int fd, ret;
78
79 fd = open(filename, O_RDONLY);
Jens Axboe4dced402007-07-23 14:16:55 +020080 if (fd < 0)
Jens Axboefb7b71a2007-05-15 08:44:04 +020081 return 0;
Jens Axboefb7b71a2007-05-15 08:44:04 +020082
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)) {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100135 dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
Jens Axboe5e6c2062007-05-16 14:40:29 +0200136 strcpy(path, full_path);
137 found = 1;
138 break;
139 }
140 }
141
142 closedir(D);
143 return found;
144}
145
Jens Axboec69aa912007-05-22 15:51:50 +0200146#define FMINORBITS 20
147#define FMINORMASK ((1U << FMINORBITS) - 1)
148#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
149#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
Jens Axboeeeb9c2a2007-05-16 18:28:47 +0200150
Jens Axboe691c8fb2008-03-07 14:26:26 +0100151static void trace_add_open_event(struct thread_data *td, int fileno)
152{
153 struct io_piece *ipo;
154
155 ipo = calloc(1, sizeof(*ipo));
156
157 ipo->ddir = DDIR_INVAL;
158 ipo->fileno = fileno;
159 ipo->file_action = FIO_LOG_OPEN_FILE;
Jens Axboe01743ee2008-06-02 12:19:19 +0200160 flist_add_tail(&ipo->list, &td->io_log_list);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100161}
162
Jens Axboe5e6c2062007-05-16 14:40:29 +0200163static void trace_add_file(struct thread_data *td, __u32 device)
164{
165 static unsigned int last_maj, last_min;
Jens Axboec69aa912007-05-22 15:51:50 +0200166 unsigned int maj = FMAJOR(device);
167 unsigned int min = FMINOR(device);
Jens Axboe5e6c2062007-05-16 14:40:29 +0200168 struct fio_file *f;
169 char dev[256];
170 unsigned int i;
171
172 if (last_maj == maj && last_min == min)
173 return;
174
175 last_maj = maj;
176 last_min = min;
177
178 /*
179 * check for this file in our list
180 */
181 for_each_file(td, f, i)
182 if (f->major == maj && f->minor == min)
183 return;
184
185 strcpy(dev, "/dev");
Jens Axboebd6f78b2008-02-01 20:27:52 +0100186 if (lookup_device(dev, maj, min)) {
Jens Axboe691c8fb2008-03-07 14:26:26 +0100187 int fileno;
188
Jens Axboebd6f78b2008-02-01 20:27:52 +0100189 dprint(FD_BLKTRACE, "add devices %s\n", dev);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100190 fileno = add_file(td, dev);
191 trace_add_open_event(td, fileno);
Jens Axboebd6f78b2008-02-01 20:27:52 +0100192 }
Jens Axboe5e6c2062007-05-16 14:40:29 +0200193}
194
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200195/*
196 * Store blk_io_trace data in an ipo for later retrieval.
197 */
Jens Axboefdefd982007-05-15 10:12:26 +0200198static void store_ipo(struct thread_data *td, unsigned long long offset,
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200199 unsigned int bytes, int rw, unsigned long long ttime)
Jens Axboefdefd982007-05-15 10:12:26 +0200200{
201 struct io_piece *ipo = malloc(sizeof(*ipo));
202
203 memset(ipo, 0, sizeof(*ipo));
Jens Axboe01743ee2008-06-02 12:19:19 +0200204 INIT_FLIST_HEAD(&ipo->list);
Jens Axboea2eea812007-05-15 13:10:41 +0200205 /*
206 * the 512 is wrong here, it should be the hardware sector size...
207 */
208 ipo->offset = offset * 512;
Jens Axboefdefd982007-05-15 10:12:26 +0200209 ipo->len = bytes;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200210 ipo->delay = ttime / 1000;
Jens Axboefdefd982007-05-15 10:12:26 +0200211 if (rw)
212 ipo->ddir = DDIR_WRITE;
213 else
214 ipo->ddir = DDIR_READ;
215
Jens Axboebd6f78b2008-02-01 20:27:52 +0100216 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
217 ipo->ddir, ipo->offset,
218 ipo->len, ipo->delay);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100219 queue_io_piece(td, ipo);
Jens Axboefdefd982007-05-15 10:12:26 +0200220}
221
Jens Axboeaec2de22008-04-24 12:44:42 +0200222static void handle_trace_notify( struct blk_io_trace *t)
Jens Axboe691c8fb2008-03-07 14:26:26 +0100223{
224 switch (t->action) {
225 case BLK_TN_PROCESS:
226 printf("got process notify: %x, %d\n", t->action, t->pid);
227 break;
228 case BLK_TN_TIMESTAMP:
229 printf("got timestamp notify: %x, %d\n", t->action, t->pid);
230 break;
231 default:
232 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
233 break;
234 }
235}
236
237static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
238 unsigned long long ttime, unsigned long *ios,
239 unsigned int *bs)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200240{
Jens Axboefdefd982007-05-15 10:12:26 +0200241 int rw;
242
Jens Axboe5e6c2062007-05-16 14:40:29 +0200243 trace_add_file(td, t->device);
244
Jens Axboee7a7d702007-05-15 10:13:04 +0200245 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
Jens Axboed84f8d42007-05-16 08:47:46 +0200246
247 if (t->bytes > bs[rw])
248 bs[rw] = t->bytes;
249
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200250 ios[rw]++;
Jens Axboe6df8ada2007-05-15 13:23:19 +0200251 td->o.size += t->bytes;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200252 store_ipo(td, t->sector, t->bytes, rw, ttime);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200253}
254
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200255/*
Jens Axboe691c8fb2008-03-07 14:26:26 +0100256 * We only care for queue traces, most of the others are side effects
257 * due to internal workings of the block layer.
258 */
259static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
260 unsigned long long ttime, unsigned long *ios,
261 unsigned int *bs)
262{
263 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
264 return;
265 if (t->action & BLK_TC_ACT(BLK_TC_PC))
266 return;
267
268 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
Jens Axboeaec2de22008-04-24 12:44:42 +0200269 handle_trace_notify(t);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100270 else
271 handle_trace_fs(td, t, ttime, ios, bs);
272}
273
274/*
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200275 * Load a blktrace file by reading all the blk_io_trace entries, and storing
276 * them as io_pieces like the fio text version would do.
277 */
Jens Axboefb7b71a2007-05-15 08:44:04 +0200278int load_blktrace(struct thread_data *td, const char *filename)
279{
Jens Axboea61edde2007-05-15 14:29:58 +0200280 unsigned long long ttime, delay;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200281 struct blk_io_trace t;
Jens Axboe4241ea82007-09-12 08:18:36 +0200282 unsigned long ios[2], skipped_writes;
Jens Axboea61edde2007-05-15 14:29:58 +0200283 unsigned int cpu;
Jens Axboed84f8d42007-05-16 08:47:46 +0200284 unsigned int rw_bs[2];
Jens Axboee2887562007-05-16 09:25:09 +0200285 struct fifo *fifo;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200286 int fd;
287
288 fd = open(filename, O_RDONLY);
289 if (fd < 0) {
290 td_verror(td, errno, "open blktrace file");
291 return 1;
292 }
293
Jens Axboee2887562007-05-16 09:25:09 +0200294 fifo = fifo_alloc(TRACE_FIFO_SIZE);
295
Jens Axboe6df8ada2007-05-15 13:23:19 +0200296 td->o.size = 0;
297
Jens Axboea61edde2007-05-15 14:29:58 +0200298 cpu = 0;
Jens Axboed84f8d42007-05-16 08:47:46 +0200299 ttime = 0;
300 ios[0] = ios[1] = 0;
301 rw_bs[0] = rw_bs[1] = 0;
Jens Axboe4241ea82007-09-12 08:18:36 +0200302 skipped_writes = 0;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200303 do {
Jens Axboee2887562007-05-16 09:25:09 +0200304 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
Jens Axboefb7b71a2007-05-15 08:44:04 +0200305
Jens Axboee2887562007-05-16 09:25:09 +0200306 if (ret < 0)
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200307 goto err;
Jens Axboee2887562007-05-16 09:25:09 +0200308 else if (!ret)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200309 break;
Jens Axboee2887562007-05-16 09:25:09 +0200310 else if (ret < (int) sizeof(t)) {
311 log_err("fio: short fifo get\n");
312 break;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200313 }
314
315 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100316 log_err("fio: bad magic in blktrace data: %x\n",
317 t.magic);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200318 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200319 }
320 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100321 log_err("fio: bad blktrace version %d\n",
322 t.magic & 0xff);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200323 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200324 }
Jens Axboef12b3232007-05-16 12:16:44 +0200325 ret = discard_pdu(td, fifo, fd, &t);
326 if (ret < 0) {
Jens Axboefb7b71a2007-05-15 08:44:04 +0200327 td_verror(td, ret, "blktrace lseek");
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200328 goto err;
Jens Axboef12b3232007-05-16 12:16:44 +0200329 } else if (t.pdu_len != ret) {
330 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
331 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200332 }
Jens Axboe691c8fb2008-03-07 14:26:26 +0100333 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
334 if (!ttime) {
335 ttime = t.time;
336 cpu = t.cpu;
337 }
338
339 delay = 0;
340 if (cpu == t.cpu)
341 delay = t.time - ttime;
342 if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
343 skipped_writes++;
344 else
345 handle_trace(td, &t, delay, ios, rw_bs);
346
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200347 ttime = t.time;
Jens Axboea61edde2007-05-15 14:29:58 +0200348 cpu = t.cpu;
Jens Axboea6edd632008-03-07 21:41:54 +0100349 } else {
350 delay = 0;
Jens Axboe4241ea82007-09-12 08:18:36 +0200351 handle_trace(td, &t, delay, ios, rw_bs);
Jens Axboea6edd632008-03-07 21:41:54 +0100352 }
Jens Axboefb7b71a2007-05-15 08:44:04 +0200353 } while (1);
354
Jens Axboe38470f82007-05-16 09:26:23 +0200355 fifo_free(fifo);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200356 close(fd);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200357
Jens Axboe4241ea82007-09-12 08:18:36 +0200358 if (skipped_writes)
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100359 log_err("fio: %s skips replay of %lu writes due to read-only\n",
360 td->o.name, skipped_writes);
Jens Axboe4241ea82007-09-12 08:18:36 +0200361
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200362 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
363 log_err("fio: found no ios in blktrace data\n");
364 return 1;
Jens Axboed84f8d42007-05-16 08:47:46 +0200365 } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200366 td->o.td_ddir = TD_DDIR_READ;
Jens Axboed84f8d42007-05-16 08:47:46 +0200367 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
368 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200369 td->o.td_ddir = TD_DDIR_WRITE;
Jens Axboed84f8d42007-05-16 08:47:46 +0200370 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
371 } else {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200372 td->o.td_ddir = TD_DDIR_RW;
Jens Axboed84f8d42007-05-16 08:47:46 +0200373 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
374 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
375 }
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200376
377 /*
378 * We need to do direct/raw ios to the device, to avoid getting
379 * read-ahead in our way.
380 */
381 td->o.odirect = 1;
382
Jens Axboefb7b71a2007-05-15 08:44:04 +0200383 return 0;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200384err:
385 close(fd);
Jens Axboe38470f82007-05-16 09:26:23 +0200386 fifo_free(fifo);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200387 return 1;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200388}