powerpc: use ATB clock, if it's available
[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;
113
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);
144
145 }
146
147 free(ipo);
148
149 if (io_u->ddir != DDIR_WAIT)
150 return 0;
151 }
152
153 td->done = 1;
154 return 1;
155}
156
157void prune_io_piece_log(struct thread_data *td)
158{
159 struct io_piece *ipo;
160 struct rb_node *n;
161
162 while ((n = rb_first(&td->io_hist_tree)) != NULL) {
163 ipo = rb_entry(n, struct io_piece, rb_node);
164 rb_erase(n, &td->io_hist_tree);
165 remove_trim_entry(td, ipo);
166 td->io_hist_len--;
167 free(ipo);
168 }
169
170 while (!flist_empty(&td->io_hist_list)) {
171 ipo = flist_entry(td->io_hist_list.next, struct io_piece, list);
172 flist_del(&ipo->list);
173 remove_trim_entry(td, ipo);
174 td->io_hist_len--;
175 free(ipo);
176 }
177}
178
179/*
180 * log a successful write, so we can unwind the log for verify
181 */
182void log_io_piece(struct thread_data *td, struct io_u *io_u)
183{
184 struct rb_node **p, *parent;
185 struct io_piece *ipo, *__ipo;
186
187 ipo = malloc(sizeof(struct io_piece));
188 init_ipo(ipo);
189 ipo->file = io_u->file;
190 ipo->offset = io_u->offset;
191 ipo->len = io_u->buflen;
192
193 if (io_u_should_trim(td, io_u)) {
194 flist_add_tail(&ipo->trim_list, &td->trim_list);
195 td->trim_entries++;
196 }
197
198 /*
199 * We don't need to sort the entries, if:
200 *
201 * Sequential writes, or
202 * Random writes that lay out the file as it goes along
203 *
204 * For both these cases, just reading back data in the order we
205 * wrote it out is the fastest.
206 *
207 * One exception is if we don't have a random map AND we are doing
208 * verifies, in that case we need to check for duplicate blocks and
209 * drop the old one, which we rely on the rb insert/lookup for
210 * handling.
211 */
212 if ((!td_random(td) || !td->o.overwrite) &&
213 (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
214 INIT_FLIST_HEAD(&ipo->list);
215 flist_add_tail(&ipo->list, &td->io_hist_list);
216 ipo->flags |= IP_F_ONLIST;
217 td->io_hist_len++;
218 return;
219 }
220
221 RB_CLEAR_NODE(&ipo->rb_node);
222
223 /*
224 * Sort the entry into the verification list
225 */
226restart:
227 p = &td->io_hist_tree.rb_node;
228 parent = NULL;
229 while (*p) {
230 parent = *p;
231
232 __ipo = rb_entry(parent, struct io_piece, rb_node);
233 if (ipo->file < __ipo->file)
234 p = &(*p)->rb_left;
235 else if (ipo->file > __ipo->file)
236 p = &(*p)->rb_right;
237 else if (ipo->offset < __ipo->offset)
238 p = &(*p)->rb_left;
239 else if (ipo->offset > __ipo->offset)
240 p = &(*p)->rb_right;
241 else {
885ac623
JA
242 dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu",
243 __ipo->offset, __ipo->len,
244 ipo->offset, ipo->len);
ac9b9101
JA
245 td->io_hist_len--;
246 rb_erase(parent, &td->io_hist_tree);
247 remove_trim_entry(td, __ipo);
248 free(__ipo);
249 goto restart;
250 }
251 }
252
253 rb_link_node(&ipo->rb_node, parent, p);
254 rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
255 ipo->flags |= IP_F_ONRB;
256 td->io_hist_len++;
257}
258
259void write_iolog_close(struct thread_data *td)
260{
261 fflush(td->iolog_f);
262 fclose(td->iolog_f);
263 free(td->iolog_buf);
264 td->iolog_f = NULL;
265 td->iolog_buf = NULL;
266}
267
268/*
269 * Read version 2 iolog data. It is enhanced to include per-file logging,
270 * syncs, etc.
271 */
272static int read_iolog2(struct thread_data *td, FILE *f)
273{
274 unsigned long long offset;
275 unsigned int bytes;
276 int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
277 char *fname, *act;
278 char *str, *p;
279 enum fio_ddir rw;
280
281 free_release_files(td);
282
283 /*
284 * Read in the read iolog and store it, reuse the infrastructure
285 * for doing verifications.
286 */
287 str = malloc(4096);
288 fname = malloc(256+16);
289 act = malloc(256+16);
290
291 reads = writes = waits = 0;
292 while ((p = fgets(str, 4096, f)) != NULL) {
293 struct io_piece *ipo;
294 int r;
295
296 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
297 &bytes);
298 if (r == 4) {
299 /*
300 * Check action first
301 */
302 if (!strcmp(act, "wait"))
303 rw = DDIR_WAIT;
304 else if (!strcmp(act, "read"))
305 rw = DDIR_READ;
306 else if (!strcmp(act, "write"))
307 rw = DDIR_WRITE;
308 else if (!strcmp(act, "sync"))
309 rw = DDIR_SYNC;
310 else if (!strcmp(act, "datasync"))
311 rw = DDIR_DATASYNC;
312 else if (!strcmp(act, "trim"))
313 rw = DDIR_TRIM;
314 else {
315 log_err("fio: bad iolog file action: %s\n",
316 act);
317 continue;
318 }
033ace1e 319 fileno = get_fileno(td, fname);
ac9b9101
JA
320 } else if (r == 2) {
321 rw = DDIR_INVAL;
322 if (!strcmp(act, "add")) {
323 td->o.nr_files++;
324 fileno = add_file(td, fname);
325 file_action = FIO_LOG_ADD_FILE;
326 continue;
327 } else if (!strcmp(act, "open")) {
328 fileno = get_fileno(td, fname);
329 file_action = FIO_LOG_OPEN_FILE;
330 } else if (!strcmp(act, "close")) {
331 fileno = get_fileno(td, fname);
332 file_action = FIO_LOG_CLOSE_FILE;
333 } else {
334 log_err("fio: bad iolog file action: %s\n",
335 act);
336 continue;
337 }
338 } else {
339 log_err("bad iolog2: %s", p);
340 continue;
341 }
342
343 if (rw == DDIR_READ)
344 reads++;
345 else if (rw == DDIR_WRITE) {
346 /*
347 * Don't add a write for ro mode
348 */
349 if (read_only)
350 continue;
351 writes++;
352 } else if (rw == DDIR_WAIT) {
353 waits++;
354 } else if (rw == DDIR_INVAL) {
355 } else if (!ddir_sync(rw)) {
356 log_err("bad ddir: %d\n", rw);
357 continue;
358 }
359
360 /*
361 * Make note of file
362 */
363 ipo = malloc(sizeof(*ipo));
364 init_ipo(ipo);
365 ipo->ddir = rw;
366 if (rw == DDIR_WAIT) {
367 ipo->delay = offset;
368 } else {
369 ipo->offset = offset;
370 ipo->len = bytes;
371 if (bytes > td->o.max_bs[rw])
372 td->o.max_bs[rw] = bytes;
373 ipo->fileno = fileno;
374 ipo->file_action = file_action;
375 }
376
377 queue_io_piece(td, ipo);
378 }
379
380 free(str);
381 free(act);
382 free(fname);
383
384 if (writes && read_only) {
385 log_err("fio: <%s> skips replay of %d writes due to"
386 " read-only\n", td->o.name, writes);
387 writes = 0;
388 }
389
390 if (!reads && !writes && !waits)
391 return 1;
392 else if (reads && !writes)
393 td->o.td_ddir = TD_DDIR_READ;
394 else if (!reads && writes)
395 td->o.td_ddir = TD_DDIR_WRITE;
396 else
397 td->o.td_ddir = TD_DDIR_RW;
398
399 return 0;
400}
401
402/*
403 * open iolog, check version, and call appropriate parser
404 */
405static int init_iolog_read(struct thread_data *td)
406{
407 char buffer[256], *p;
408 FILE *f;
409 int ret;
410
411 f = fopen(td->o.read_iolog_file, "r");
412 if (!f) {
413 perror("fopen read iolog");
414 return 1;
415 }
416
417 p = fgets(buffer, sizeof(buffer), f);
418 if (!p) {
419 td_verror(td, errno, "iolog read");
420 log_err("fio: unable to read iolog\n");
421 fclose(f);
422 return 1;
423 }
424
425 /*
426 * version 2 of the iolog stores a specific string as the
427 * first line, check for that
428 */
429 if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
430 ret = read_iolog2(td, f);
431 else {
432 log_err("fio: iolog version 1 is no longer supported\n");
433 ret = 1;
434 }
435
436 fclose(f);
437 return ret;
438}
439
440/*
441 * Set up a log for storing io patterns.
442 */
443static int init_iolog_write(struct thread_data *td)
444{
445 struct fio_file *ff;
446 FILE *f;
447 unsigned int i;
448
449 f = fopen(td->o.write_iolog_file, "a");
450 if (!f) {
451 perror("fopen write iolog");
452 return 1;
453 }
454
455 /*
456 * That's it for writing, setup a log buffer and we're done.
457 */
458 td->iolog_f = f;
459 td->iolog_buf = malloc(8192);
460 setvbuf(f, td->iolog_buf, _IOFBF, 8192);
461
462 /*
463 * write our version line
464 */
465 if (fprintf(f, "%s\n", iolog_ver2) < 0) {
466 perror("iolog init\n");
467 return 1;
468 }
469
470 /*
471 * add all known files
472 */
473 for_each_file(td, ff, i)
474 log_file(td, ff, FIO_LOG_ADD_FILE);
475
476 return 0;
477}
478
479int init_iolog(struct thread_data *td)
480{
481 int ret = 0;
482
483 if (td->o.read_iolog_file) {
484 /*
485 * Check if it's a blktrace file and load that if possible.
486 * Otherwise assume it's a normal log file and load that.
487 */
488 if (is_blktrace(td->o.read_iolog_file))
489 ret = load_blktrace(td, td->o.read_iolog_file);
490 else
491 ret = init_iolog_read(td);
492 } else if (td->o.write_iolog_file)
493 ret = init_iolog_write(td);
494
495 return ret;
496}
497
b8bc8cba 498void setup_log(struct io_log **log, unsigned long avg_msec)
ac9b9101
JA
499{
500 struct io_log *l = malloc(sizeof(*l));
501
b8bc8cba 502 memset(l, 0, sizeof(*l));
ac9b9101
JA
503 l->nr_samples = 0;
504 l->max_samples = 1024;
505 l->log = malloc(l->max_samples * sizeof(struct io_sample));
b8bc8cba 506 l->avg_msec = avg_msec;
ac9b9101
JA
507 *log = l;
508}
509
510void __finish_log(struct io_log *log, const char *name)
511{
512 unsigned int i;
513 FILE *f;
514
515 f = fopen(name, "a");
516 if (!f) {
517 perror("fopen log");
518 return;
519 }
520
521 for (i = 0; i < log->nr_samples; i++) {
522 fprintf(f, "%lu, %lu, %u, %u\n", log->log[i].time,
523 log->log[i].val,
524 log->log[i].ddir,
525 log->log[i].bs);
526 }
527
528 fclose(f);
529 free(log->log);
530 free(log);
531}
532
533void finish_log_named(struct thread_data *td, struct io_log *log,
534 const char *prefix, const char *postfix)
535{
536 char file_name[256], *p;
537
98ffb8f3 538 snprintf(file_name, sizeof(file_name), "%s_%s.log", prefix, postfix);
ac9b9101
JA
539 p = basename(file_name);
540 __finish_log(log, p);
541}
542
543void finish_log(struct thread_data *td, struct io_log *log, const char *name)
544{
545 finish_log_named(td, log, td->o.name, name);
546}