blob: c44d8221ee936c4c0d235c87efa0f7607eca8dc7 [file] [log] [blame]
Jens Axboef29b25a2007-07-23 08:56:43 +02001/*
2 * Code related to writing an iolog of what a thread is doing, and to
3 * later read that back and replay
4 */
Jens Axboe3c39a372006-06-06 20:56:12 +02005#include <stdio.h>
6#include <stdlib.h>
Jens Axboe5921e802008-05-30 15:02:38 +02007#include <libgen.h>
Jens Axboef29b25a2007-07-23 08:56:43 +02008#include <assert.h>
Jens Axboe01743ee2008-06-02 12:19:19 +02009#include "flist.h"
Jens Axboe3c39a372006-06-06 20:56:12 +020010#include "fio.h"
Jens Axboe4f5af7b2009-06-03 08:45:40 +020011#include "verify.h"
Jens Axboe3c39a372006-06-06 20:56:12 +020012
Jens Axboef29b25a2007-07-23 08:56:43 +020013static const char iolog_ver2[] = "fio version 2 iolog";
14
Jens Axboe691c8fb2008-03-07 14:26:26 +010015void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
16{
Jens Axboe01743ee2008-06-02 12:19:19 +020017 flist_add_tail(&ipo->list, &td->io_log_list);
Jens Axboe691c8fb2008-03-07 14:26:26 +010018 td->total_io_size += ipo->len;
19}
20
Jens Axboef29b25a2007-07-23 08:56:43 +020021void log_io_u(struct thread_data *td, struct io_u *io_u)
Jens Axboe3c39a372006-06-06 20:56:12 +020022{
Jens Axboe5f9099e2009-06-16 22:40:26 +020023 const char *act[] = { "read", "write", "sync", "datasync" };
Jens Axboef29b25a2007-07-23 08:56:43 +020024
25 assert(io_u->ddir < 3);
26
27 if (!td->o.write_iolog_file)
28 return;
29
Jens Axboe5ec10ea2008-03-06 15:42:00 +010030 fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
31 act[io_u->ddir], io_u->offset,
32 io_u->buflen);
Jens Axboef29b25a2007-07-23 08:56:43 +020033}
34
35void log_file(struct thread_data *td, struct fio_file *f,
36 enum file_log_act what)
37{
38 const char *act[] = { "add", "open", "close" };
39
40 assert(what < 3);
41
42 if (!td->o.write_iolog_file)
43 return;
44
Jens Axboe393ca7e2008-05-15 09:17:42 +020045
46 /*
47 * this happens on the pre-open/close done before the job starts
48 */
49 if (!td->iolog_f)
50 return;
51
Jens Axboef29b25a2007-07-23 08:56:43 +020052 fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
Jens Axboe3c39a372006-06-06 20:56:12 +020053}
54
Jens Axboea61edde2007-05-15 14:29:58 +020055static void iolog_delay(struct thread_data *td, unsigned long delay)
56{
57 unsigned long usec = utime_since_now(&td->last_issue);
58
59 if (delay < usec)
60 return;
61
62 delay -= usec;
63
64 /*
65 * less than 100 usec delay, just regard it as noise
66 */
67 if (delay < 100)
68 return;
69
70 usec_sleep(td, delay);
71}
72
Jens Axboef7182732008-03-10 13:57:58 +010073static int ipo_special(struct thread_data *td, struct io_piece *ipo)
74{
75 struct fio_file *f;
76 int ret;
77
78 /*
79 * Not a special ipo
80 */
81 if (ipo->ddir != DDIR_INVAL)
82 return 0;
83
84 f = td->files[ipo->fileno];
85
86 switch (ipo->file_action) {
87 case FIO_LOG_OPEN_FILE:
88 ret = td_io_open_file(td, f);
89 if (!ret) {
90 free(ipo);
91 break;
92 }
93 td_verror(td, ret, "iolog open file");
94 return -1;
95 case FIO_LOG_CLOSE_FILE:
96 td_io_close_file(td, f);
97 break;
98 case FIO_LOG_UNLINK_FILE:
99 unlink(f->file_name);
100 break;
101 default:
102 log_err("fio: bad file action %d\n", ipo->file_action);
103 break;
104 }
105
106 return 1;
107}
108
Jens Axboe3c39a372006-06-06 20:56:12 +0200109int read_iolog_get(struct thread_data *td, struct io_u *io_u)
110{
111 struct io_piece *ipo;
112
Jens Axboe01743ee2008-06-02 12:19:19 +0200113 while (!flist_empty(&td->io_log_list)) {
Jens Axboef7182732008-03-10 13:57:58 +0100114 int ret;
115
Jens Axboe01743ee2008-06-02 12:19:19 +0200116 ipo = flist_entry(td->io_log_list.next, struct io_piece, list);
117 flist_del(&ipo->list);
Jens Axboea61edde2007-05-15 14:29:58 +0200118
Jens Axboef7182732008-03-10 13:57:58 +0100119 ret = ipo_special(td, ipo);
120 if (ret < 0) {
121 free(ipo);
122 break;
123 } else if (ret > 0) {
124 free(ipo);
125 continue;
Jens Axboef29b25a2007-07-23 08:56:43 +0200126 }
127
Jens Axboe429f6672007-07-23 10:38:43 +0200128 io_u->offset = ipo->offset;
129 io_u->buflen = ipo->len;
130 io_u->ddir = ipo->ddir;
Jens Axboe126d65c2008-03-01 18:04:31 +0100131 io_u->file = td->files[ipo->fileno];
Jens Axboe429f6672007-07-23 10:38:43 +0200132 get_file(io_u->file);
133
Jens Axboeee56ad52008-02-01 10:30:20 +0100134 dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
135 io_u->buflen, io_u->file->file_name);
136
Jens Axboea61edde2007-05-15 14:29:58 +0200137 if (ipo->delay)
138 iolog_delay(td, ipo->delay);
139
Jens Axboe3c39a372006-06-06 20:56:12 +0200140 free(ipo);
141 return 0;
142 }
143
Jens Axboe20e354e2007-07-23 14:36:16 +0200144 td->done = 1;
Jens Axboe3c39a372006-06-06 20:56:12 +0200145 return 1;
146}
147
148void prune_io_piece_log(struct thread_data *td)
149{
150 struct io_piece *ipo;
Jens Axboe4b878982007-03-26 09:32:22 +0200151 struct rb_node *n;
Jens Axboe3c39a372006-06-06 20:56:12 +0200152
Jens Axboe4b878982007-03-26 09:32:22 +0200153 while ((n = rb_first(&td->io_hist_tree)) != NULL) {
154 ipo = rb_entry(n, struct io_piece, rb_node);
155 rb_erase(n, &td->io_hist_tree);
Jens Axboe3c39a372006-06-06 20:56:12 +0200156 free(ipo);
157 }
Jens Axboe8ce9cd32008-02-21 09:52:35 +0100158
Jens Axboe01743ee2008-06-02 12:19:19 +0200159 while (!flist_empty(&td->io_hist_list)) {
160 ipo = flist_entry(td->io_hist_list.next, struct io_piece, list);
161 flist_del(&ipo->list);
Jens Axboe8ce9cd32008-02-21 09:52:35 +0100162 free(ipo);
163 }
Jens Axboe3c39a372006-06-06 20:56:12 +0200164}
165
166/*
Jens Axboe34403fb2007-03-02 21:43:25 +0100167 * log a successful write, so we can unwind the log for verify
Jens Axboe3c39a372006-06-06 20:56:12 +0200168 */
169void log_io_piece(struct thread_data *td, struct io_u *io_u)
170{
Jens Axboe8de8f042007-03-27 10:36:12 +0200171 struct rb_node **p, *parent;
Jens Axboe4b878982007-03-26 09:32:22 +0200172 struct io_piece *ipo, *__ipo;
Jens Axboe3c39a372006-06-06 20:56:12 +0200173
Jens Axboe4b878982007-03-26 09:32:22 +0200174 ipo = malloc(sizeof(struct io_piece));
Jens Axboe53cdc682006-10-18 11:50:58 +0200175 ipo->file = io_u->file;
Jens Axboe3c39a372006-06-06 20:56:12 +0200176 ipo->offset = io_u->offset;
177 ipo->len = io_u->buflen;
178
179 /*
Jens Axboe8de8f042007-03-27 10:36:12 +0200180 * We don't need to sort the entries, if:
181 *
182 * Sequential writes, or
183 * Random writes that lay out the file as it goes along
184 *
185 * For both these cases, just reading back data in the order we
186 * wrote it out is the fastest.
Jens Axboe83472392009-02-19 21:32:12 +0100187 *
188 * One exception is if we don't have a random map AND we are doing
189 * verifies, in that case we need to check for duplicate blocks and
190 * drop the old one, which we rely on the rb insert/lookup for
191 * handling.
Jens Axboe8de8f042007-03-27 10:36:12 +0200192 */
Jens Axboe83472392009-02-19 21:32:12 +0100193 if ((!td_random(td) || !td->o.overwrite) &&
194 (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
Jens Axboe01743ee2008-06-02 12:19:19 +0200195 INIT_FLIST_HEAD(&ipo->list);
196 flist_add_tail(&ipo->list, &td->io_hist_list);
Jens Axboe8de8f042007-03-27 10:36:12 +0200197 return;
198 }
199
200 RB_CLEAR_NODE(&ipo->rb_node);
Jens Axboe8de8f042007-03-27 10:36:12 +0200201
202 /*
Jens Axboe4b878982007-03-26 09:32:22 +0200203 * Sort the entry into the verification list
Jens Axboe3c39a372006-06-06 20:56:12 +0200204 */
Jens Axboe83472392009-02-19 21:32:12 +0100205restart:
206 p = &td->io_hist_tree.rb_node;
207 parent = NULL;
Jens Axboe4b878982007-03-26 09:32:22 +0200208 while (*p) {
209 parent = *p;
Jens Axboe3c39a372006-06-06 20:56:12 +0200210
Jens Axboe4b878982007-03-26 09:32:22 +0200211 __ipo = rb_entry(parent, struct io_piece, rb_node);
Jens Axboe83472392009-02-19 21:32:12 +0100212 if (ipo->offset < __ipo->offset)
Jens Axboe4b878982007-03-26 09:32:22 +0200213 p = &(*p)->rb_left;
Jens Axboe83472392009-02-19 21:32:12 +0100214 else if (ipo->offset > __ipo->offset)
Jens Axboebb5d7d02007-03-27 10:21:25 +0200215 p = &(*p)->rb_right;
Jens Axboe83472392009-02-19 21:32:12 +0100216 else {
217 assert(ipo->len == __ipo->len);
218 rb_erase(parent, &td->io_hist_tree);
219 goto restart;
220 }
Jens Axboe3c39a372006-06-06 20:56:12 +0200221 }
222
Jens Axboe4b878982007-03-26 09:32:22 +0200223 rb_link_node(&ipo->rb_node, parent, p);
224 rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
Jens Axboe3c39a372006-06-06 20:56:12 +0200225}
226
227void write_iolog_close(struct thread_data *td)
228{
229 fflush(td->iolog_f);
230 fclose(td->iolog_f);
231 free(td->iolog_buf);
Jens Axboef29b25a2007-07-23 08:56:43 +0200232 td->iolog_f = NULL;
233 td->iolog_buf = NULL;
Jens Axboe3c39a372006-06-06 20:56:12 +0200234}
235
Jens Axboefb71fbd2006-10-20 09:15:46 +0200236/*
Jens Axboef29b25a2007-07-23 08:56:43 +0200237 * Read version 2 iolog data. It is enhanced to include per-file logging,
238 * syncs, etc.
Jens Axboefb71fbd2006-10-20 09:15:46 +0200239 */
Jens Axboef29b25a2007-07-23 08:56:43 +0200240static int read_iolog2(struct thread_data *td, FILE *f)
241{
242 unsigned long long offset;
243 unsigned int bytes;
Jens Axboe53fa9b62007-07-23 11:25:39 +0200244 int reads, writes, fileno = 0, file_action = 0; /* stupid gcc */
Jens Axboef29b25a2007-07-23 08:56:43 +0200245 char *fname, *act;
246 char *str, *p;
Jens Axboe53fa9b62007-07-23 11:25:39 +0200247 enum fio_ddir rw;
Jens Axboef29b25a2007-07-23 08:56:43 +0200248
249 free_release_files(td);
250
251 /*
252 * Read in the read iolog and store it, reuse the infrastructure
253 * for doing verifications.
254 */
255 str = malloc(4096);
256 fname = malloc(256+16);
257 act = malloc(256+16);
258
259 reads = writes = 0;
260 while ((p = fgets(str, 4096, f)) != NULL) {
261 struct io_piece *ipo;
262 int r;
263
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100264 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
265 &bytes);
Jens Axboef29b25a2007-07-23 08:56:43 +0200266 if (r == 4) {
267 /*
268 * Check action first
269 */
270 if (!strcmp(act, "read"))
271 rw = DDIR_READ;
272 else if (!strcmp(act, "write"))
273 rw = DDIR_WRITE;
274 else if (!strcmp(act, "sync"))
275 rw = DDIR_SYNC;
Jens Axboe5f9099e2009-06-16 22:40:26 +0200276 else if (!strcmp(act, "datasync"))
277 rw = DDIR_DATASYNC;
Jens Axboef29b25a2007-07-23 08:56:43 +0200278 else {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100279 log_err("fio: bad iolog file action: %s\n",
280 act);
Jens Axboef29b25a2007-07-23 08:56:43 +0200281 continue;
282 }
283 } else if (r == 2) {
284 rw = DDIR_INVAL;
285 if (!strcmp(act, "add")) {
286 td->o.nr_files++;
287 fileno = add_file(td, fname);
288 file_action = FIO_LOG_ADD_FILE;
289 continue;
290 } else if (!strcmp(act, "open")) {
291 fileno = get_fileno(td, fname);
292 file_action = FIO_LOG_OPEN_FILE;
293 } else if (!strcmp(act, "close")) {
294 fileno = get_fileno(td, fname);
295 file_action = FIO_LOG_CLOSE_FILE;
296 } else {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100297 log_err("fio: bad iolog file action: %s\n",
298 act);
Jens Axboef29b25a2007-07-23 08:56:43 +0200299 continue;
300 }
301 } else {
302 log_err("bad iolog2: %s", p);
303 continue;
304 }
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100305
Jens Axboef29b25a2007-07-23 08:56:43 +0200306 if (rw == DDIR_READ)
307 reads++;
Jens Axboe4241ea82007-09-12 08:18:36 +0200308 else if (rw == DDIR_WRITE) {
Jens Axboe4241ea82007-09-12 08:18:36 +0200309 /*
310 * Don't add a write for ro mode
311 */
312 if (read_only)
313 continue;
Jens Axboeed4aa702008-03-07 13:39:59 +0100314 writes++;
Jens Axboe5f9099e2009-06-16 22:40:26 +0200315 } else if (!ddir_sync(rw)) {
Jens Axboef29b25a2007-07-23 08:56:43 +0200316 log_err("bad ddir: %d\n", rw);
317 continue;
318 }
319
320 /*
321 * Make note of file
322 */
323 ipo = malloc(sizeof(*ipo));
324 memset(ipo, 0, sizeof(*ipo));
Jens Axboe01743ee2008-06-02 12:19:19 +0200325 INIT_FLIST_HEAD(&ipo->list);
Jens Axboef29b25a2007-07-23 08:56:43 +0200326 ipo->offset = offset;
327 ipo->len = bytes;
Jens Axboe53fa9b62007-07-23 11:25:39 +0200328 ipo->ddir = rw;
Jens Axboef29b25a2007-07-23 08:56:43 +0200329 if (bytes > td->o.max_bs[rw])
330 td->o.max_bs[rw] = bytes;
331 if (rw == DDIR_INVAL) {
332 ipo->fileno = fileno;
333 ipo->file_action = file_action;
334 }
Jens Axboe691c8fb2008-03-07 14:26:26 +0100335 queue_io_piece(td, ipo);
Jens Axboef29b25a2007-07-23 08:56:43 +0200336 }
337
338 free(str);
339 free(act);
340 free(fname);
341
Jens Axboe4241ea82007-09-12 08:18:36 +0200342 if (writes && read_only) {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100343 log_err("fio: <%s> skips replay of %d writes due to"
344 " read-only\n", td->o.name, writes);
Jens Axboe4241ea82007-09-12 08:18:36 +0200345 writes = 0;
346 }
347
Jens Axboef29b25a2007-07-23 08:56:43 +0200348 if (!reads && !writes)
349 return 1;
350 else if (reads && !writes)
351 td->o.td_ddir = TD_DDIR_READ;
352 else if (!reads && writes)
353 td->o.td_ddir = TD_DDIR_WRITE;
354 else
355 td->o.td_ddir = TD_DDIR_RW;
356
357 return 0;
358}
359
360/*
Jens Axboef29b25a2007-07-23 08:56:43 +0200361 * open iolog, check version, and call appropriate parser
362 */
363static int init_iolog_read(struct thread_data *td)
364{
365 char buffer[256], *p;
366 FILE *f;
367 int ret;
368
369 f = fopen(td->o.read_iolog_file, "r");
370 if (!f) {
371 perror("fopen read iolog");
372 return 1;
373 }
374
375 p = fgets(buffer, sizeof(buffer), f);
376 if (!p) {
377 td_verror(td, errno, "iolog read");
378 log_err("fio: unable to read iolog\n");
379 return 1;
380 }
381
382 /*
383 * version 2 of the iolog stores a specific string as the
384 * first line, check for that
385 */
386 if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
387 ret = read_iolog2(td, f);
388 else {
Jens Axboeaec2de22008-04-24 12:44:42 +0200389 log_err("fio: iolog version 1 is no longer supported\n");
390 ret = 1;
Jens Axboef29b25a2007-07-23 08:56:43 +0200391 }
392
393 fclose(f);
394 return ret;
395}
396
397/*
Jens Axboefb71fbd2006-10-20 09:15:46 +0200398 * Setup a log for storing io patterns.
399 */
400static int init_iolog_write(struct thread_data *td)
401{
Jens Axboef29b25a2007-07-23 08:56:43 +0200402 struct fio_file *ff;
Jens Axboe076efc72006-10-27 11:24:25 +0200403 FILE *f;
Jens Axboef29b25a2007-07-23 08:56:43 +0200404 unsigned int i;
Jens Axboe733ed592007-04-25 14:24:12 +0200405
Jens Axboec12f6da2008-11-20 09:13:04 +0100406 f = fopen(td->o.write_iolog_file, "a");
Jens Axboefb71fbd2006-10-20 09:15:46 +0200407 if (!f) {
408 perror("fopen write iolog");
409 return 1;
410 }
411
412 /*
413 * That's it for writing, setup a log buffer and we're done.
414 */
415 td->iolog_f = f;
416 td->iolog_buf = malloc(8192);
417 setvbuf(f, td->iolog_buf, _IOFBF, 8192);
Jens Axboef29b25a2007-07-23 08:56:43 +0200418
419 /*
420 * write our version line
421 */
422 if (fprintf(f, "%s\n", iolog_ver2) < 0) {
423 perror("iolog init\n");
424 return 1;
425 }
426
427 /*
428 * add all known files
429 */
430 for_each_file(td, ff, i)
431 log_file(td, ff, FIO_LOG_ADD_FILE);
432
Jens Axboefb71fbd2006-10-20 09:15:46 +0200433 return 0;
434}
435
436int init_iolog(struct thread_data *td)
437{
Jens Axboeb4a6a592006-10-20 13:54:47 +0200438 int ret = 0;
Jens Axboefb71fbd2006-10-20 09:15:46 +0200439
Jens Axboefb7b71a2007-05-15 08:44:04 +0200440 if (td->o.read_iolog_file) {
441 /*
442 * Check if it's a blktrace file and load that if possible.
443 * Otherwise assume it's a normal log file and load that.
444 */
445 if (is_blktrace(td->o.read_iolog_file))
446 ret = load_blktrace(td, td->o.read_iolog_file);
447 else
448 ret = init_iolog_read(td);
449 } else if (td->o.write_iolog_file)
Jens Axboeb4a6a592006-10-20 13:54:47 +0200450 ret = init_iolog_write(td);
451
Jens Axboe1e97cce2006-12-05 11:44:16 +0100452 return ret;
Jens Axboefb71fbd2006-10-20 09:15:46 +0200453}
454
Jens Axboe8914a9d2006-06-07 11:14:56 +0200455void setup_log(struct io_log **log)
456{
457 struct io_log *l = malloc(sizeof(*l));
458
459 l->nr_samples = 0;
460 l->max_samples = 1024;
461 l->log = malloc(l->max_samples * sizeof(struct io_sample));
462 *log = l;
463}
464
Jens Axboebb3884d2007-01-17 17:23:11 +1100465void __finish_log(struct io_log *log, const char *name)
Jens Axboe8914a9d2006-06-07 11:14:56 +0200466{
Jens Axboe8914a9d2006-06-07 11:14:56 +0200467 unsigned int i;
Jens Axboebb3884d2007-01-17 17:23:11 +1100468 FILE *f;
Jens Axboe8914a9d2006-06-07 11:14:56 +0200469
Jens Axboec12f6da2008-11-20 09:13:04 +0100470 f = fopen(name, "a");
Jens Axboe8914a9d2006-06-07 11:14:56 +0200471 if (!f) {
472 perror("fopen log");
473 return;
474 }
475
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100476 for (i = 0; i < log->nr_samples; i++) {
Jens Axboe306ddc92009-05-18 13:08:12 +0200477 fprintf(f, "%lu, %lu, %u, %u\n", log->log[i].time,
478 log->log[i].val,
479 log->log[i].ddir,
480 log->log[i].bs);
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100481 }
Jens Axboe8914a9d2006-06-07 11:14:56 +0200482
483 fclose(f);
484 free(log->log);
485 free(log);
486}
Jens Axboebb3884d2007-01-17 17:23:11 +1100487
Jens Axboee3cedca2008-11-19 19:57:52 +0100488void finish_log_named(struct thread_data *td, struct io_log *log,
489 const char *prefix, const char *postfix)
Jens Axboebb3884d2007-01-17 17:23:11 +1100490{
Jens Axboe748b23a2008-05-07 14:28:22 +0200491 char file_name[256], *p;
Jens Axboebb3884d2007-01-17 17:23:11 +1100492
Jens Axboee3cedca2008-11-19 19:57:52 +0100493 snprintf(file_name, 200, "%s_%s.log", prefix, postfix);
Jens Axboe748b23a2008-05-07 14:28:22 +0200494 p = basename(file_name);
495 __finish_log(log, p);
Jens Axboebb3884d2007-01-17 17:23:11 +1100496}
Jens Axboee3cedca2008-11-19 19:57:52 +0100497
498void finish_log(struct thread_data *td, struct io_log *log, const char *name)
499{
500 finish_log_named(td, log, td->o.name, name);
501}