blktrace: add support for non-native endian format
[fio.git] / iolog.c
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
14 static const char iolog_ver2[] = "fio version 2 iolog";
15
16 void 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
22 void 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
37 void 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
57 static 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
75 static 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
109 int 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                 free(ipo);
147
148                 if (io_u->ddir != DDIR_WAIT)
149                         return 0;
150         }
151
152         td->done = 1;
153         return 1;
154 }
155
156 void 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  */
181 void 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;
191
192         if (io_u_should_trim(td, io_u)) {
193                 flist_add_tail(&ipo->trim_list, &td->trim_list);
194                 td->trim_entries++;
195         }
196
197         /*
198          * We don't need to sort the entries, if:
199          *
200          *      Sequential writes, or
201          *      Random writes that lay out the file as it goes along
202          *
203          * For both these cases, just reading back data in the order we
204          * wrote it out is the fastest.
205          *
206          * One exception is if we don't have a random map AND we are doing
207          * verifies, in that case we need to check for duplicate blocks and
208          * drop the old one, which we rely on the rb insert/lookup for
209          * handling.
210          */
211         if ((!td_random(td) || !td->o.overwrite) &&
212               (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
213                 INIT_FLIST_HEAD(&ipo->list);
214                 flist_add_tail(&ipo->list, &td->io_hist_list);
215                 ipo->flags |= IP_F_ONLIST;
216                 td->io_hist_len++;
217                 return;
218         }
219
220         RB_CLEAR_NODE(&ipo->rb_node);
221
222         /*
223          * Sort the entry into the verification list
224          */
225 restart:
226         p = &td->io_hist_tree.rb_node;
227         parent = NULL;
228         while (*p) {
229                 parent = *p;
230
231                 __ipo = rb_entry(parent, struct io_piece, rb_node);
232                 if (ipo->file < __ipo->file)
233                         p = &(*p)->rb_left;
234                 else if (ipo->file > __ipo->file)
235                         p = &(*p)->rb_right;
236                 else if (ipo->offset < __ipo->offset)
237                         p = &(*p)->rb_left;
238                 else if (ipo->offset > __ipo->offset)
239                         p = &(*p)->rb_right;
240                 else {
241                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu",
242                                 __ipo->offset, __ipo->len,
243                                 ipo->offset, ipo->len);
244                         td->io_hist_len--;
245                         rb_erase(parent, &td->io_hist_tree);
246                         remove_trim_entry(td, __ipo);
247                         free(__ipo);
248                         goto restart;
249                 }
250         }
251
252         rb_link_node(&ipo->rb_node, parent, p);
253         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
254         ipo->flags |= IP_F_ONRB;
255         td->io_hist_len++;
256 }
257
258 void write_iolog_close(struct thread_data *td)
259 {
260         fflush(td->iolog_f);
261         fclose(td->iolog_f);
262         free(td->iolog_buf);
263         td->iolog_f = NULL;
264         td->iolog_buf = NULL;
265 }
266
267 /*
268  * Read version 2 iolog data. It is enhanced to include per-file logging,
269  * syncs, etc.
270  */
271 static int read_iolog2(struct thread_data *td, FILE *f)
272 {
273         unsigned long long offset;
274         unsigned int bytes;
275         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
276         char *fname, *act;
277         char *str, *p;
278         enum fio_ddir rw;
279
280         free_release_files(td);
281
282         /*
283          * Read in the read iolog and store it, reuse the infrastructure
284          * for doing verifications.
285          */
286         str = malloc(4096);
287         fname = malloc(256+16);
288         act = malloc(256+16);
289
290         reads = writes = waits = 0;
291         while ((p = fgets(str, 4096, f)) != NULL) {
292                 struct io_piece *ipo;
293                 int r;
294
295                 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
296                                                                         &bytes);
297                 if (r == 4) {
298                         /*
299                          * Check action first
300                          */
301                         if (!strcmp(act, "wait"))
302                                 rw = DDIR_WAIT;
303                         else if (!strcmp(act, "read"))
304                                 rw = DDIR_READ;
305                         else if (!strcmp(act, "write"))
306                                 rw = DDIR_WRITE;
307                         else if (!strcmp(act, "sync"))
308                                 rw = DDIR_SYNC;
309                         else if (!strcmp(act, "datasync"))
310                                 rw = DDIR_DATASYNC;
311                         else if (!strcmp(act, "trim"))
312                                 rw = DDIR_TRIM;
313                         else {
314                                 log_err("fio: bad iolog file action: %s\n",
315                                                                         act);
316                                 continue;
317                         }
318                         fileno = get_fileno(td, fname);
319                 } else if (r == 2) {
320                         rw = DDIR_INVAL;
321                         if (!strcmp(act, "add")) {
322                                 td->o.nr_files++;
323                                 fileno = add_file(td, fname);
324                                 file_action = FIO_LOG_ADD_FILE;
325                                 continue;
326                         } else if (!strcmp(act, "open")) {
327                                 fileno = get_fileno(td, fname);
328                                 file_action = FIO_LOG_OPEN_FILE;
329                         } else if (!strcmp(act, "close")) {
330                                 fileno = get_fileno(td, fname);
331                                 file_action = FIO_LOG_CLOSE_FILE;
332                         } else {
333                                 log_err("fio: bad iolog file action: %s\n",
334                                                                         act);
335                                 continue;
336                         }
337                 } else {
338                         log_err("bad iolog2: %s", p);
339                         continue;
340                 }
341
342                 if (rw == DDIR_READ)
343                         reads++;
344                 else if (rw == DDIR_WRITE) {
345                         /*
346                          * Don't add a write for ro mode
347                          */
348                         if (read_only)
349                                 continue;
350                         writes++;
351                 } else if (rw == DDIR_WAIT) {
352                         waits++;
353                 } else if (rw == DDIR_INVAL) {
354                 } else if (!ddir_sync(rw)) {
355                         log_err("bad ddir: %d\n", rw);
356                         continue;
357                 }
358
359                 /*
360                  * Make note of file
361                  */
362                 ipo = malloc(sizeof(*ipo));
363                 init_ipo(ipo);
364                 ipo->ddir = rw;
365                 if (rw == DDIR_WAIT) {
366                         ipo->delay = offset;
367                 } else {
368                         ipo->offset = offset;
369                         ipo->len = bytes;
370                         if (bytes > td->o.max_bs[rw])
371                                 td->o.max_bs[rw] = bytes;
372                         ipo->fileno = fileno;
373                         ipo->file_action = file_action;
374                 }
375
376                 queue_io_piece(td, ipo);
377         }
378
379         free(str);
380         free(act);
381         free(fname);
382
383         if (writes && read_only) {
384                 log_err("fio: <%s> skips replay of %d writes due to"
385                         " read-only\n", td->o.name, writes);
386                 writes = 0;
387         }
388
389         if (!reads && !writes && !waits)
390                 return 1;
391         else if (reads && !writes)
392                 td->o.td_ddir = TD_DDIR_READ;
393         else if (!reads && writes)
394                 td->o.td_ddir = TD_DDIR_WRITE;
395         else
396                 td->o.td_ddir = TD_DDIR_RW;
397
398         return 0;
399 }
400
401 /*
402  * open iolog, check version, and call appropriate parser
403  */
404 static int init_iolog_read(struct thread_data *td)
405 {
406         char buffer[256], *p;
407         FILE *f;
408         int ret;
409
410         f = fopen(td->o.read_iolog_file, "r");
411         if (!f) {
412                 perror("fopen read iolog");
413                 return 1;
414         }
415
416         p = fgets(buffer, sizeof(buffer), f);
417         if (!p) {
418                 td_verror(td, errno, "iolog read");
419                 log_err("fio: unable to read iolog\n");
420                 fclose(f);
421                 return 1;
422         }
423
424         /*
425          * version 2 of the iolog stores a specific string as the
426          * first line, check for that
427          */
428         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
429                 ret = read_iolog2(td, f);
430         else {
431                 log_err("fio: iolog version 1 is no longer supported\n");
432                 ret = 1;
433         }
434
435         fclose(f);
436         return ret;
437 }
438
439 /*
440  * Set up a log for storing io patterns.
441  */
442 static int init_iolog_write(struct thread_data *td)
443 {
444         struct fio_file *ff;
445         FILE *f;
446         unsigned int i;
447
448         f = fopen(td->o.write_iolog_file, "a");
449         if (!f) {
450                 perror("fopen write iolog");
451                 return 1;
452         }
453
454         /*
455          * That's it for writing, setup a log buffer and we're done.
456           */
457         td->iolog_f = f;
458         td->iolog_buf = malloc(8192);
459         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
460
461         /*
462          * write our version line
463          */
464         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
465                 perror("iolog init\n");
466                 return 1;
467         }
468
469         /*
470          * add all known files
471          */
472         for_each_file(td, ff, i)
473                 log_file(td, ff, FIO_LOG_ADD_FILE);
474
475         return 0;
476 }
477
478 int init_iolog(struct thread_data *td)
479 {
480         int ret = 0;
481
482         if (td->o.read_iolog_file) {
483                 int need_swap;
484
485                 /*
486                  * Check if it's a blktrace file and load that if possible.
487                  * Otherwise assume it's a normal log file and load that.
488                  */
489                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
490                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
491                 else
492                         ret = init_iolog_read(td);
493         } else if (td->o.write_iolog_file)
494                 ret = init_iolog_write(td);
495
496         return ret;
497 }
498
499 void setup_log(struct io_log **log, unsigned long avg_msec, int log_type)
500 {
501         struct io_log *l = malloc(sizeof(*l));
502
503         memset(l, 0, sizeof(*l));
504         l->nr_samples = 0;
505         l->max_samples = 1024;
506         l->log_type = log_type;
507         l->log = malloc(l->max_samples * sizeof(struct io_sample));
508         l->avg_msec = avg_msec;
509         *log = l;
510 }
511
512 void __finish_log(struct io_log *log, const char *name)
513 {
514         unsigned int i;
515         FILE *f;
516
517         f = fopen(name, "a");
518         if (!f) {
519                 perror("fopen log");
520                 return;
521         }
522
523         for (i = 0; i < log->nr_samples; i++) {
524                 fprintf(f, "%lu, %lu, %u, %u\n",
525                                 (unsigned long) log->log[i].time,
526                                 (unsigned long) log->log[i].val,
527                                 log->log[i].ddir, log->log[i].bs);
528         }
529
530         fclose(f);
531         free(log->log);
532         free(log);
533 }
534
535 void finish_log_named(struct thread_data *td, struct io_log *log,
536                        const char *prefix, const char *postfix)
537 {
538         char file_name[256], *p;
539
540         snprintf(file_name, sizeof(file_name), "%s_%s.log", prefix, postfix);
541         p = basename(file_name);
542
543         if (td->client_type == FIO_CLIENT_TYPE_GUI) {
544                 fio_send_iolog(td, log, p);
545                 free(log->log);
546                 free(log);
547         } else
548                 __finish_log(log, p);
549 }
550
551 void finish_log(struct thread_data *td, struct io_log *log, const char *name)
552 {
553         finish_log_named(td, log, td->o.name, name);
554 }