verify: always log IO in the order they are issued
[fio.git] / iolog.c
CommitLineData
ac9b9101
JA
1/*
2 * Code related to writing an iolog of what a thread is doing, and to
3 * later read that back and replay
4 */
5#include <stdio.h>
6#include <stdlib.h>
7#include <libgen.h>
8#include <assert.h>
9#include "flist.h"
10#include "fio.h"
11#include "verify.h"
12#include "trim.h"
13
14static const char iolog_ver2[] = "fio version 2 iolog";
15
16void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
17{
18 flist_add_tail(&ipo->list, &td->io_log_list);
19 td->total_io_size += ipo->len;
20}
21
22void log_io_u(struct thread_data *td, struct io_u *io_u)
23{
24 const char *act[] = { "read", "write", "sync", "datasync",
25 "sync_file_range", "wait", "trim" };
26
27 assert(io_u->ddir <= 6);
28
29 if (!td->o.write_iolog_file)
30 return;
31
32 fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
33 act[io_u->ddir], io_u->offset,
34 io_u->buflen);
35}
36
37void log_file(struct thread_data *td, struct fio_file *f,
38 enum file_log_act what)
39{
40 const char *act[] = { "add", "open", "close" };
41
42 assert(what < 3);
43
44 if (!td->o.write_iolog_file)
45 return;
46
47
48 /*
49 * this happens on the pre-open/close done before the job starts
50 */
51 if (!td->iolog_f)
52 return;
53
54 fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
55}
56
57static void iolog_delay(struct thread_data *td, unsigned long delay)
58{
59 unsigned long usec = utime_since_now(&td->last_issue);
60
61 if (delay < usec)
62 return;
63
64 delay -= usec;
65
66 /*
67 * less than 100 usec delay, just regard it as noise
68 */
69 if (delay < 100)
70 return;
71
72 usec_sleep(td, delay);
73}
74
75static int ipo_special(struct thread_data *td, struct io_piece *ipo)
76{
77 struct fio_file *f;
78 int ret;
79
80 /*
81 * Not a special ipo
82 */
83 if (ipo->ddir != DDIR_INVAL)
84 return 0;
85
86 f = td->files[ipo->fileno];
87
88 switch (ipo->file_action) {
89 case FIO_LOG_OPEN_FILE:
90 ret = td_io_open_file(td, f);
91 if (!ret)
92 break;
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
109int read_iolog_get(struct thread_data *td, struct io_u *io_u)
110{
111 struct io_piece *ipo;
112 unsigned long elapsed;
3c3ed070 113
ac9b9101
JA
114 while (!flist_empty(&td->io_log_list)) {
115 int ret;
116
117 ipo = flist_entry(td->io_log_list.next, struct io_piece, list);
118 flist_del(&ipo->list);
119 remove_trim_entry(td, ipo);
120
121 ret = ipo_special(td, ipo);
122 if (ret < 0) {
123 free(ipo);
124 break;
125 } else if (ret > 0) {
126 free(ipo);
127 continue;
128 }
129
130 io_u->ddir = ipo->ddir;
131 if (ipo->ddir != DDIR_WAIT) {
132 io_u->offset = ipo->offset;
133 io_u->buflen = ipo->len;
134 io_u->file = td->files[ipo->fileno];
135 get_file(io_u->file);
136 dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
137 io_u->buflen, io_u->file->file_name);
138 if (ipo->delay)
139 iolog_delay(td, ipo->delay);
140 } else {
141 elapsed = mtime_since_genesis();
142 if (ipo->delay > elapsed)
143 usec_sleep(td, (ipo->delay - elapsed) * 1000);
ac9b9101
JA
144 }
145
146 free(ipo);
3c3ed070 147
ac9b9101
JA
148 if (io_u->ddir != DDIR_WAIT)
149 return 0;
150 }
151
152 td->done = 1;
153 return 1;
154}
155
156void prune_io_piece_log(struct thread_data *td)
157{
158 struct io_piece *ipo;
159 struct rb_node *n;
160
161 while ((n = rb_first(&td->io_hist_tree)) != NULL) {
162 ipo = rb_entry(n, struct io_piece, rb_node);
163 rb_erase(n, &td->io_hist_tree);
164 remove_trim_entry(td, ipo);
165 td->io_hist_len--;
166 free(ipo);
167 }
168
169 while (!flist_empty(&td->io_hist_list)) {
170 ipo = flist_entry(td->io_hist_list.next, struct io_piece, list);
171 flist_del(&ipo->list);
172 remove_trim_entry(td, ipo);
173 td->io_hist_len--;
174 free(ipo);
175 }
176}
177
178/*
179 * log a successful write, so we can unwind the log for verify
180 */
181void log_io_piece(struct thread_data *td, struct io_u *io_u)
182{
183 struct rb_node **p, *parent;
184 struct io_piece *ipo, *__ipo;
185
186 ipo = malloc(sizeof(struct io_piece));
187 init_ipo(ipo);
188 ipo->file = io_u->file;
189 ipo->offset = io_u->offset;
190 ipo->len = io_u->buflen;
da0a7bd2 191 ipo->numberio = io_u->numberio;
f9401285
JA
192 ipo->flags = IP_F_IN_FLIGHT;
193
194 io_u->ipo = ipo;
ac9b9101
JA
195
196 if (io_u_should_trim(td, io_u)) {
197 flist_add_tail(&ipo->trim_list, &td->trim_list);
198 td->trim_entries++;
199 }
200
201 /*
202 * We don't need to sort the entries, if:
203 *
204 * Sequential writes, or
205 * Random writes that lay out the file as it goes along
206 *
207 * For both these cases, just reading back data in the order we
208 * wrote it out is the fastest.
209 *
210 * One exception is if we don't have a random map AND we are doing
211 * verifies, in that case we need to check for duplicate blocks and
212 * drop the old one, which we rely on the rb insert/lookup for
213 * handling.
214 */
c4b6117b 215 if (((!td->o.verifysort) || !td_random(td) || !td->o.overwrite) &&
ac9b9101
JA
216 (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
217 INIT_FLIST_HEAD(&ipo->list);
218 flist_add_tail(&ipo->list, &td->io_hist_list);
219 ipo->flags |= IP_F_ONLIST;
220 td->io_hist_len++;
221 return;
222 }
223
224 RB_CLEAR_NODE(&ipo->rb_node);
225
226 /*
227 * Sort the entry into the verification list
228 */
229restart:
230 p = &td->io_hist_tree.rb_node;
231 parent = NULL;
232 while (*p) {
233 parent = *p;
234
235 __ipo = rb_entry(parent, struct io_piece, rb_node);
236 if (ipo->file < __ipo->file)
237 p = &(*p)->rb_left;
238 else if (ipo->file > __ipo->file)
239 p = &(*p)->rb_right;
240 else if (ipo->offset < __ipo->offset)
241 p = &(*p)->rb_left;
242 else if (ipo->offset > __ipo->offset)
243 p = &(*p)->rb_right;
244 else {
885ac623
JA
245 dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu",
246 __ipo->offset, __ipo->len,
247 ipo->offset, ipo->len);
ac9b9101
JA
248 td->io_hist_len--;
249 rb_erase(parent, &td->io_hist_tree);
250 remove_trim_entry(td, __ipo);
251 free(__ipo);
252 goto restart;
253 }
254 }
255
256 rb_link_node(&ipo->rb_node, parent, p);
257 rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
258 ipo->flags |= IP_F_ONRB;
259 td->io_hist_len++;
260}
261
262void write_iolog_close(struct thread_data *td)
263{
264 fflush(td->iolog_f);
265 fclose(td->iolog_f);
266 free(td->iolog_buf);
267 td->iolog_f = NULL;
268 td->iolog_buf = NULL;
269}
270
271/*
272 * Read version 2 iolog data. It is enhanced to include per-file logging,
273 * syncs, etc.
274 */
275static int read_iolog2(struct thread_data *td, FILE *f)
276{
277 unsigned long long offset;
278 unsigned int bytes;
279 int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
280 char *fname, *act;
281 char *str, *p;
282 enum fio_ddir rw;
283
284 free_release_files(td);
285
286 /*
287 * Read in the read iolog and store it, reuse the infrastructure
288 * for doing verifications.
289 */
290 str = malloc(4096);
291 fname = malloc(256+16);
292 act = malloc(256+16);
293
294 reads = writes = waits = 0;
295 while ((p = fgets(str, 4096, f)) != NULL) {
296 struct io_piece *ipo;
297 int r;
298
299 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
300 &bytes);
301 if (r == 4) {
302 /*
303 * Check action first
304 */
305 if (!strcmp(act, "wait"))
306 rw = DDIR_WAIT;
307 else if (!strcmp(act, "read"))
308 rw = DDIR_READ;
309 else if (!strcmp(act, "write"))
310 rw = DDIR_WRITE;
311 else if (!strcmp(act, "sync"))
312 rw = DDIR_SYNC;
313 else if (!strcmp(act, "datasync"))
314 rw = DDIR_DATASYNC;
315 else if (!strcmp(act, "trim"))
316 rw = DDIR_TRIM;
317 else {
318 log_err("fio: bad iolog file action: %s\n",
319 act);
320 continue;
321 }
033ace1e 322 fileno = get_fileno(td, fname);
ac9b9101
JA
323 } else if (r == 2) {
324 rw = DDIR_INVAL;
325 if (!strcmp(act, "add")) {
326 td->o.nr_files++;
327 fileno = add_file(td, fname);
328 file_action = FIO_LOG_ADD_FILE;
329 continue;
330 } else if (!strcmp(act, "open")) {
331 fileno = get_fileno(td, fname);
332 file_action = FIO_LOG_OPEN_FILE;
333 } else if (!strcmp(act, "close")) {
334 fileno = get_fileno(td, fname);
335 file_action = FIO_LOG_CLOSE_FILE;
336 } else {
337 log_err("fio: bad iolog file action: %s\n",
338 act);
339 continue;
340 }
341 } else {
342 log_err("bad iolog2: %s", p);
343 continue;
344 }
345
346 if (rw == DDIR_READ)
347 reads++;
348 else if (rw == DDIR_WRITE) {
349 /*
350 * Don't add a write for ro mode
351 */
352 if (read_only)
353 continue;
354 writes++;
355 } else if (rw == DDIR_WAIT) {
356 waits++;
357 } else if (rw == DDIR_INVAL) {
358 } else if (!ddir_sync(rw)) {
359 log_err("bad ddir: %d\n", rw);
360 continue;
361 }
362
363 /*
364 * Make note of file
365 */
366 ipo = malloc(sizeof(*ipo));
367 init_ipo(ipo);
368 ipo->ddir = rw;
369 if (rw == DDIR_WAIT) {
370 ipo->delay = offset;
371 } else {
372 ipo->offset = offset;
373 ipo->len = bytes;
374 if (bytes > td->o.max_bs[rw])
375 td->o.max_bs[rw] = bytes;
376 ipo->fileno = fileno;
377 ipo->file_action = file_action;
378 }
3c3ed070 379
ac9b9101
JA
380 queue_io_piece(td, ipo);
381 }
382
383 free(str);
384 free(act);
385 free(fname);
386
387 if (writes && read_only) {
388 log_err("fio: <%s> skips replay of %d writes due to"
389 " read-only\n", td->o.name, writes);
390 writes = 0;
391 }
392
393 if (!reads && !writes && !waits)
394 return 1;
395 else if (reads && !writes)
396 td->o.td_ddir = TD_DDIR_READ;
397 else if (!reads && writes)
398 td->o.td_ddir = TD_DDIR_WRITE;
399 else
400 td->o.td_ddir = TD_DDIR_RW;
401
402 return 0;
403}
404
405/*
406 * open iolog, check version, and call appropriate parser
407 */
408static int init_iolog_read(struct thread_data *td)
409{
410 char buffer[256], *p;
411 FILE *f;
412 int ret;
413
414 f = fopen(td->o.read_iolog_file, "r");
415 if (!f) {
416 perror("fopen read iolog");
417 return 1;
418 }
419
420 p = fgets(buffer, sizeof(buffer), f);
421 if (!p) {
422 td_verror(td, errno, "iolog read");
423 log_err("fio: unable to read iolog\n");
424 fclose(f);
425 return 1;
426 }
427
428 /*
429 * version 2 of the iolog stores a specific string as the
430 * first line, check for that
431 */
432 if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
433 ret = read_iolog2(td, f);
434 else {
435 log_err("fio: iolog version 1 is no longer supported\n");
436 ret = 1;
437 }
438
439 fclose(f);
440 return ret;
441}
442
443/*
444 * Set up a log for storing io patterns.
445 */
446static int init_iolog_write(struct thread_data *td)
447{
448 struct fio_file *ff;
449 FILE *f;
450 unsigned int i;
451
452 f = fopen(td->o.write_iolog_file, "a");
453 if (!f) {
454 perror("fopen write iolog");
455 return 1;
456 }
457
458 /*
459 * That's it for writing, setup a log buffer and we're done.
460 */
461 td->iolog_f = f;
462 td->iolog_buf = malloc(8192);
463 setvbuf(f, td->iolog_buf, _IOFBF, 8192);
464
465 /*
466 * write our version line
467 */
468 if (fprintf(f, "%s\n", iolog_ver2) < 0) {
469 perror("iolog init\n");
470 return 1;
471 }
472
473 /*
474 * add all known files
475 */
476 for_each_file(td, ff, i)
477 log_file(td, ff, FIO_LOG_ADD_FILE);
478
479 return 0;
480}
481
482int init_iolog(struct thread_data *td)
483{
484 int ret = 0;
485
486 if (td->o.read_iolog_file) {
d95b34a6
JA
487 int need_swap;
488
ac9b9101
JA
489 /*
490 * Check if it's a blktrace file and load that if possible.
491 * Otherwise assume it's a normal log file and load that.
492 */
d95b34a6
JA
493 if (is_blktrace(td->o.read_iolog_file, &need_swap))
494 ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
ac9b9101
JA
495 else
496 ret = init_iolog_read(td);
497 } else if (td->o.write_iolog_file)
498 ret = init_iolog_write(td);
499
f01b34ae
JA
500 if (ret)
501 td_verror(td, EINVAL, "failed initializing iolog");
502
ac9b9101
JA
503 return ret;
504}
505
ea51b956 506void setup_log(struct io_log **log, unsigned long avg_msec, int log_type)
ac9b9101
JA
507{
508 struct io_log *l = malloc(sizeof(*l));
509
b8bc8cba 510 memset(l, 0, sizeof(*l));
ac9b9101
JA
511 l->nr_samples = 0;
512 l->max_samples = 1024;
ea51b956 513 l->log_type = log_type;
ac9b9101 514 l->log = malloc(l->max_samples * sizeof(struct io_sample));
b8bc8cba 515 l->avg_msec = avg_msec;
ac9b9101
JA
516 *log = l;
517}
518
519void __finish_log(struct io_log *log, const char *name)
520{
521 unsigned int i;
522 FILE *f;
523
524 f = fopen(name, "a");
525 if (!f) {
526 perror("fopen log");
527 return;
528 }
529
530 for (i = 0; i < log->nr_samples; i++) {
1fae4855
JA
531 fprintf(f, "%lu, %lu, %u, %u\n",
532 (unsigned long) log->log[i].time,
533 (unsigned long) log->log[i].val,
534 log->log[i].ddir, log->log[i].bs);
ac9b9101
JA
535 }
536
537 fclose(f);
538 free(log->log);
539 free(log);
540}
541
542void finish_log_named(struct thread_data *td, struct io_log *log,
543 const char *prefix, const char *postfix)
544{
545 char file_name[256], *p;
546
98ffb8f3 547 snprintf(file_name, sizeof(file_name), "%s_%s.log", prefix, postfix);
ac9b9101 548 p = basename(file_name);
1b42725f 549
f5ed765a 550 if (td->client_type == FIO_CLIENT_TYPE_GUI) {
1b42725f 551 fio_send_iolog(td, log, p);
f5ed765a
JA
552 free(log->log);
553 free(log);
554 } else
555 __finish_log(log, p);
ac9b9101
JA
556}
557
558void finish_log(struct thread_data *td, struct io_log *log, const char *name)
559{
560 finish_log_named(td, log, td->o.name, name);
561}