blob: 297a8a947af254200aed536788b095811a0d34ac [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 Axboe2da7df12010-08-01 21:27:28 +020013#define TRACE_FIFO_SIZE 8192
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
David Nellansd1c46c02010-08-31 21:20:47 +0200100static int lookup_device(struct thread_data *td, char *path, unsigned int maj,
101 unsigned int min)
Jens Axboe5e6c2062007-05-16 14:40:29 +0200102{
103 struct dirent *dir;
104 struct stat st;
105 int found = 0;
106 DIR *D;
107
108 D = opendir(path);
109 if (!D)
110 return 0;
111
112 while ((dir = readdir(D)) != NULL) {
113 char full_path[256];
114
115 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
116 continue;
117
118 sprintf(full_path, "%s/%s", path, dir->d_name);
119 if (lstat(full_path, &st) == -1) {
120 perror("lstat");
121 break;
122 }
123
124 if (S_ISDIR(st.st_mode)) {
David Nellansd1c46c02010-08-31 21:20:47 +0200125 found = lookup_device(td, full_path, maj, min);
Jens Axboe5e6c2062007-05-16 14:40:29 +0200126 if (found) {
127 strcpy(path, full_path);
128 break;
129 }
130 }
131
132 if (!S_ISBLK(st.st_mode))
133 continue;
134
David Nellansd1c46c02010-08-31 21:20:47 +0200135 /*
136 * If replay_redirect is set then always return this device
137 * upon lookup which overrides the device lookup based on
138 * major minor in the actual blktrace
139 */
140 if (td->o.replay_redirect) {
141 dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
142 " with: %s", maj, min,
143 td->o.replay_redirect);
144 strcpy(path, td->o.replay_redirect);
145 found = 1;
146 break;
147 }
148
Jens Axboe5e6c2062007-05-16 14:40:29 +0200149 if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100150 dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
Jens Axboe5e6c2062007-05-16 14:40:29 +0200151 strcpy(path, full_path);
152 found = 1;
153 break;
154 }
155 }
156
157 closedir(D);
158 return found;
159}
160
Jens Axboec69aa912007-05-22 15:51:50 +0200161#define FMINORBITS 20
162#define FMINORMASK ((1U << FMINORBITS) - 1)
163#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
164#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
Jens Axboeeeb9c2a2007-05-16 18:28:47 +0200165
Jens Axboe691c8fb2008-03-07 14:26:26 +0100166static void trace_add_open_event(struct thread_data *td, int fileno)
167{
168 struct io_piece *ipo;
169
170 ipo = calloc(1, sizeof(*ipo));
Jens Axboe0d29de82010-09-01 13:54:15 +0200171 init_ipo(ipo);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100172
173 ipo->ddir = DDIR_INVAL;
174 ipo->fileno = fileno;
175 ipo->file_action = FIO_LOG_OPEN_FILE;
Jens Axboe01743ee2008-06-02 12:19:19 +0200176 flist_add_tail(&ipo->list, &td->io_log_list);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100177}
178
Jens Axboe5e6c2062007-05-16 14:40:29 +0200179static void trace_add_file(struct thread_data *td, __u32 device)
180{
181 static unsigned int last_maj, last_min;
Jens Axboec69aa912007-05-22 15:51:50 +0200182 unsigned int maj = FMAJOR(device);
183 unsigned int min = FMINOR(device);
Jens Axboe5e6c2062007-05-16 14:40:29 +0200184 struct fio_file *f;
185 char dev[256];
186 unsigned int i;
187
188 if (last_maj == maj && last_min == min)
189 return;
190
191 last_maj = maj;
192 last_min = min;
193
194 /*
195 * check for this file in our list
196 */
197 for_each_file(td, f, i)
198 if (f->major == maj && f->minor == min)
199 return;
200
201 strcpy(dev, "/dev");
David Nellansd1c46c02010-08-31 21:20:47 +0200202 if (lookup_device(td, dev, maj, min)) {
Jens Axboe691c8fb2008-03-07 14:26:26 +0100203 int fileno;
204
Jens Axboebd6f78b2008-02-01 20:27:52 +0100205 dprint(FD_BLKTRACE, "add devices %s\n", dev);
Jens Axboe49ffb4a2010-08-24 15:01:37 +0200206 fileno = add_file_exclusive(td, dev);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100207 trace_add_open_event(td, fileno);
Jens Axboebd6f78b2008-02-01 20:27:52 +0100208 }
Jens Axboe5e6c2062007-05-16 14:40:29 +0200209}
210
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200211/*
212 * Store blk_io_trace data in an ipo for later retrieval.
213 */
Jens Axboefdefd982007-05-15 10:12:26 +0200214static void store_ipo(struct thread_data *td, unsigned long long offset,
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200215 unsigned int bytes, int rw, unsigned long long ttime)
Jens Axboefdefd982007-05-15 10:12:26 +0200216{
217 struct io_piece *ipo = malloc(sizeof(*ipo));
218
Jens Axboe0d29de82010-09-01 13:54:15 +0200219 init_ipo(ipo);
220
Jens Axboea2eea812007-05-15 13:10:41 +0200221 /*
222 * the 512 is wrong here, it should be the hardware sector size...
223 */
224 ipo->offset = offset * 512;
Jens Axboefdefd982007-05-15 10:12:26 +0200225 ipo->len = bytes;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200226 ipo->delay = ttime / 1000;
Jens Axboefdefd982007-05-15 10:12:26 +0200227 if (rw)
228 ipo->ddir = DDIR_WRITE;
229 else
230 ipo->ddir = DDIR_READ;
231
Jens Axboebd6f78b2008-02-01 20:27:52 +0100232 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
233 ipo->ddir, ipo->offset,
234 ipo->len, ipo->delay);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100235 queue_io_piece(td, ipo);
Jens Axboefdefd982007-05-15 10:12:26 +0200236}
237
Jens Axboe0b9d69e2009-09-11 22:29:54 +0200238static void handle_trace_notify(struct blk_io_trace *t)
Jens Axboe691c8fb2008-03-07 14:26:26 +0100239{
240 switch (t->action) {
241 case BLK_TN_PROCESS:
242 printf("got process notify: %x, %d\n", t->action, t->pid);
243 break;
244 case BLK_TN_TIMESTAMP:
245 printf("got timestamp notify: %x, %d\n", t->action, t->pid);
246 break;
Jens Axboeff58fce2010-08-25 12:02:08 +0200247 case BLK_TN_MESSAGE:
248 break;
Jens Axboe691c8fb2008-03-07 14:26:26 +0100249 default:
250 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
251 break;
252 }
253}
254
Jens Axboeff58fce2010-08-25 12:02:08 +0200255static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t,
256 unsigned long long ttime, unsigned long *ios)
257{
258 struct io_piece *ipo = malloc(sizeof(*ipo));
259
Jens Axboe0d29de82010-09-01 13:54:15 +0200260 init_ipo(ipo);
Jens Axboeff58fce2010-08-25 12:02:08 +0200261 trace_add_file(td, t->device);
262
263 ios[DDIR_WRITE]++;
264 td->o.size += t->bytes;
265
266 memset(ipo, 0, sizeof(*ipo));
267 INIT_FLIST_HEAD(&ipo->list);
268
269 /*
270 * the 512 is wrong here, it should be the hardware sector size...
271 */
272 ipo->offset = t->sector * 512;
273 ipo->len = t->bytes;
274 ipo->delay = ttime / 1000;
275 ipo->ddir = DDIR_TRIM;
276
277 dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
278 ipo->offset, ipo->len,
279 ipo->delay);
280 queue_io_piece(td, ipo);
281}
282
Jens Axboe691c8fb2008-03-07 14:26:26 +0100283static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
284 unsigned long long ttime, unsigned long *ios,
285 unsigned int *bs)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200286{
Jens Axboefdefd982007-05-15 10:12:26 +0200287 int rw;
288
Jens Axboe5e6c2062007-05-16 14:40:29 +0200289 trace_add_file(td, t->device);
290
Jens Axboee7a7d702007-05-15 10:13:04 +0200291 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
Jens Axboed84f8d42007-05-16 08:47:46 +0200292
293 if (t->bytes > bs[rw])
294 bs[rw] = t->bytes;
295
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200296 ios[rw]++;
Jens Axboe6df8ada2007-05-15 13:23:19 +0200297 td->o.size += t->bytes;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200298 store_ipo(td, t->sector, t->bytes, rw, ttime);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200299}
300
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200301/*
Jens Axboe691c8fb2008-03-07 14:26:26 +0100302 * We only care for queue traces, most of the others are side effects
303 * due to internal workings of the block layer.
304 */
305static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
306 unsigned long long ttime, unsigned long *ios,
307 unsigned int *bs)
308{
309 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
310 return;
311 if (t->action & BLK_TC_ACT(BLK_TC_PC))
312 return;
313
314 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
Jens Axboeaec2de22008-04-24 12:44:42 +0200315 handle_trace_notify(t);
Jens Axboeff58fce2010-08-25 12:02:08 +0200316 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
317 handle_trace_discard(td, t, ttime, ios);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100318 else
319 handle_trace_fs(td, t, ttime, ios, bs);
320}
321
322/*
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200323 * Load a blktrace file by reading all the blk_io_trace entries, and storing
324 * them as io_pieces like the fio text version would do.
325 */
Jens Axboefb7b71a2007-05-15 08:44:04 +0200326int load_blktrace(struct thread_data *td, const char *filename)
327{
Jens Axboea61edde2007-05-15 14:29:58 +0200328 unsigned long long ttime, delay;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200329 struct blk_io_trace t;
Jens Axboe4241ea82007-09-12 08:18:36 +0200330 unsigned long ios[2], skipped_writes;
Jens Axboea61edde2007-05-15 14:29:58 +0200331 unsigned int cpu;
Jens Axboed84f8d42007-05-16 08:47:46 +0200332 unsigned int rw_bs[2];
Jens Axboee2887562007-05-16 09:25:09 +0200333 struct fifo *fifo;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200334 int fd;
335
336 fd = open(filename, O_RDONLY);
337 if (fd < 0) {
338 td_verror(td, errno, "open blktrace file");
339 return 1;
340 }
341
Jens Axboee2887562007-05-16 09:25:09 +0200342 fifo = fifo_alloc(TRACE_FIFO_SIZE);
343
Jens Axboe6df8ada2007-05-15 13:23:19 +0200344 td->o.size = 0;
345
Jens Axboea61edde2007-05-15 14:29:58 +0200346 cpu = 0;
Jens Axboed84f8d42007-05-16 08:47:46 +0200347 ttime = 0;
348 ios[0] = ios[1] = 0;
349 rw_bs[0] = rw_bs[1] = 0;
Jens Axboe4241ea82007-09-12 08:18:36 +0200350 skipped_writes = 0;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200351 do {
Jens Axboee2887562007-05-16 09:25:09 +0200352 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
Jens Axboefb7b71a2007-05-15 08:44:04 +0200353
Jens Axboee2887562007-05-16 09:25:09 +0200354 if (ret < 0)
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200355 goto err;
Jens Axboee2887562007-05-16 09:25:09 +0200356 else if (!ret)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200357 break;
Jens Axboee2887562007-05-16 09:25:09 +0200358 else if (ret < (int) sizeof(t)) {
359 log_err("fio: short fifo get\n");
360 break;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200361 }
362
363 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100364 log_err("fio: bad magic in blktrace data: %x\n",
365 t.magic);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200366 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200367 }
368 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100369 log_err("fio: bad blktrace version %d\n",
370 t.magic & 0xff);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200371 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200372 }
Jens Axboef12b3232007-05-16 12:16:44 +0200373 ret = discard_pdu(td, fifo, fd, &t);
374 if (ret < 0) {
Jens Axboefb7b71a2007-05-15 08:44:04 +0200375 td_verror(td, ret, "blktrace lseek");
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200376 goto err;
Jens Axboef12b3232007-05-16 12:16:44 +0200377 } else if (t.pdu_len != ret) {
378 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
379 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200380 }
Jens Axboe691c8fb2008-03-07 14:26:26 +0100381 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
382 if (!ttime) {
383 ttime = t.time;
384 cpu = t.cpu;
385 }
386
387 delay = 0;
388 if (cpu == t.cpu)
389 delay = t.time - ttime;
390 if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
391 skipped_writes++;
David Nellans64bbb862010-08-24 22:13:30 +0200392 else {
393 /*
394 * set delay to zero if no_stall enabled for
395 * fast replay
396 */
397 if (td->o.no_stall)
398 delay = 0;
399
Jens Axboe691c8fb2008-03-07 14:26:26 +0100400 handle_trace(td, &t, delay, ios, rw_bs);
David Nellans64bbb862010-08-24 22:13:30 +0200401 }
Jens Axboe691c8fb2008-03-07 14:26:26 +0100402
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200403 ttime = t.time;
Jens Axboea61edde2007-05-15 14:29:58 +0200404 cpu = t.cpu;
Jens Axboea6edd632008-03-07 21:41:54 +0100405 } else {
406 delay = 0;
Jens Axboe4241ea82007-09-12 08:18:36 +0200407 handle_trace(td, &t, delay, ios, rw_bs);
Jens Axboea6edd632008-03-07 21:41:54 +0100408 }
Jens Axboefb7b71a2007-05-15 08:44:04 +0200409 } while (1);
410
Jens Axboe38470f82007-05-16 09:26:23 +0200411 fifo_free(fifo);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200412 close(fd);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200413
Jens Axboe4241ea82007-09-12 08:18:36 +0200414 if (skipped_writes)
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100415 log_err("fio: %s skips replay of %lu writes due to read-only\n",
416 td->o.name, skipped_writes);
Jens Axboe4241ea82007-09-12 08:18:36 +0200417
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200418 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
419 log_err("fio: found no ios in blktrace data\n");
420 return 1;
Jens Axboed84f8d42007-05-16 08:47:46 +0200421 } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200422 td->o.td_ddir = TD_DDIR_READ;
Jens Axboed84f8d42007-05-16 08:47:46 +0200423 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
424 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200425 td->o.td_ddir = TD_DDIR_WRITE;
Jens Axboed84f8d42007-05-16 08:47:46 +0200426 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
427 } else {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200428 td->o.td_ddir = TD_DDIR_RW;
Jens Axboed84f8d42007-05-16 08:47:46 +0200429 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
430 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
431 }
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200432
433 /*
434 * We need to do direct/raw ios to the device, to avoid getting
435 * read-ahead in our way.
436 */
437 td->o.odirect = 1;
438
Jens Axboefb7b71a2007-05-15 08:44:04 +0200439 return 0;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200440err:
441 close(fd);
Jens Axboe38470f82007-05-16 09:26:23 +0200442 fifo_free(fifo);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200443 return 1;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200444}