Option for changing log files to use Unix epoch instead of being
[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 <sys/types.h>
10 #include <sys/stat.h>
11 #include <unistd.h>
12 #ifdef CONFIG_ZLIB
13 #include <zlib.h>
14 #endif
15
16 #include "flist.h"
17 #include "fio.h"
18 #include "verify.h"
19 #include "trim.h"
20 #include "filelock.h"
21 #include "smalloc.h"
22
23 static int iolog_flush(struct io_log *log);
24
25 static const char iolog_ver2[] = "fio version 2 iolog";
26
27 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
28 {
29         flist_add_tail(&ipo->list, &td->io_log_list);
30         td->total_io_size += ipo->len;
31 }
32
33 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
34 {
35         if (!td->o.write_iolog_file)
36                 return;
37
38         fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
39                                                 io_ddir_name(io_u->ddir),
40                                                 io_u->offset, io_u->buflen);
41 }
42
43 void log_file(struct thread_data *td, struct fio_file *f,
44               enum file_log_act what)
45 {
46         const char *act[] = { "add", "open", "close" };
47
48         assert(what < 3);
49
50         if (!td->o.write_iolog_file)
51                 return;
52
53
54         /*
55          * this happens on the pre-open/close done before the job starts
56          */
57         if (!td->iolog_f)
58                 return;
59
60         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
61 }
62
63 static void iolog_delay(struct thread_data *td, unsigned long delay)
64 {
65         uint64_t usec = utime_since_now(&td->last_issue);
66         uint64_t this_delay;
67         struct timeval tv;
68
69         if (delay < td->time_offset) {
70                 td->time_offset = 0;
71                 return;
72         }
73
74         delay -= td->time_offset;
75         if (delay < usec)
76                 return;
77
78         delay -= usec;
79
80         fio_gettime(&tv, NULL);
81         while (delay && !td->terminate) {
82                 this_delay = delay;
83                 if (this_delay > 500000)
84                         this_delay = 500000;
85
86                 usec_sleep(td, this_delay);
87                 delay -= this_delay;
88         }
89
90         usec = utime_since_now(&tv);
91         if (usec > delay)
92                 td->time_offset = usec - delay;
93         else
94                 td->time_offset = 0;
95 }
96
97 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
98 {
99         struct fio_file *f;
100         int ret;
101
102         /*
103          * Not a special ipo
104          */
105         if (ipo->ddir != DDIR_INVAL)
106                 return 0;
107
108         f = td->files[ipo->fileno];
109
110         switch (ipo->file_action) {
111         case FIO_LOG_OPEN_FILE:
112                 ret = td_io_open_file(td, f);
113                 if (!ret)
114                         break;
115                 td_verror(td, ret, "iolog open file");
116                 return -1;
117         case FIO_LOG_CLOSE_FILE:
118                 td_io_close_file(td, f);
119                 break;
120         case FIO_LOG_UNLINK_FILE:
121                 td_io_unlink_file(td, f);
122                 break;
123         default:
124                 log_err("fio: bad file action %d\n", ipo->file_action);
125                 break;
126         }
127
128         return 1;
129 }
130
131 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
132 {
133         struct io_piece *ipo;
134         unsigned long elapsed;
135
136         while (!flist_empty(&td->io_log_list)) {
137                 int ret;
138
139                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
140                 flist_del(&ipo->list);
141                 remove_trim_entry(td, ipo);
142
143                 ret = ipo_special(td, ipo);
144                 if (ret < 0) {
145                         free(ipo);
146                         break;
147                 } else if (ret > 0) {
148                         free(ipo);
149                         continue;
150                 }
151
152                 io_u->ddir = ipo->ddir;
153                 if (ipo->ddir != DDIR_WAIT) {
154                         io_u->offset = ipo->offset;
155                         io_u->buflen = ipo->len;
156                         io_u->file = td->files[ipo->fileno];
157                         get_file(io_u->file);
158                         dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
159                                                 io_u->buflen, io_u->file->file_name);
160                         if (ipo->delay)
161                                 iolog_delay(td, ipo->delay);
162                 } else {
163                         elapsed = mtime_since_genesis();
164                         if (ipo->delay > elapsed)
165                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
166                 }
167
168                 free(ipo);
169
170                 if (io_u->ddir != DDIR_WAIT)
171                         return 0;
172         }
173
174         td->done = 1;
175         return 1;
176 }
177
178 void prune_io_piece_log(struct thread_data *td)
179 {
180         struct io_piece *ipo;
181         struct rb_node *n;
182
183         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
184                 ipo = rb_entry(n, struct io_piece, rb_node);
185                 rb_erase(n, &td->io_hist_tree);
186                 remove_trim_entry(td, ipo);
187                 td->io_hist_len--;
188                 free(ipo);
189         }
190
191         while (!flist_empty(&td->io_hist_list)) {
192                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
193                 flist_del(&ipo->list);
194                 remove_trim_entry(td, ipo);
195                 td->io_hist_len--;
196                 free(ipo);
197         }
198 }
199
200 /*
201  * log a successful write, so we can unwind the log for verify
202  */
203 void log_io_piece(struct thread_data *td, struct io_u *io_u)
204 {
205         struct rb_node **p, *parent;
206         struct io_piece *ipo, *__ipo;
207
208         ipo = malloc(sizeof(struct io_piece));
209         init_ipo(ipo);
210         ipo->file = io_u->file;
211         ipo->offset = io_u->offset;
212         ipo->len = io_u->buflen;
213         ipo->numberio = io_u->numberio;
214         ipo->flags = IP_F_IN_FLIGHT;
215
216         io_u->ipo = ipo;
217
218         if (io_u_should_trim(td, io_u)) {
219                 flist_add_tail(&ipo->trim_list, &td->trim_list);
220                 td->trim_entries++;
221         }
222
223         /*
224          * We don't need to sort the entries, if:
225          *
226          *      Sequential writes, or
227          *      Random writes that lay out the file as it goes along
228          *
229          * For both these cases, just reading back data in the order we
230          * wrote it out is the fastest.
231          *
232          * One exception is if we don't have a random map AND we are doing
233          * verifies, in that case we need to check for duplicate blocks and
234          * drop the old one, which we rely on the rb insert/lookup for
235          * handling.
236          */
237         if (((!td->o.verifysort) || !td_random(td) || !td->o.overwrite) &&
238               (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
239                 INIT_FLIST_HEAD(&ipo->list);
240                 flist_add_tail(&ipo->list, &td->io_hist_list);
241                 ipo->flags |= IP_F_ONLIST;
242                 td->io_hist_len++;
243                 return;
244         }
245
246         RB_CLEAR_NODE(&ipo->rb_node);
247
248         /*
249          * Sort the entry into the verification list
250          */
251 restart:
252         p = &td->io_hist_tree.rb_node;
253         parent = NULL;
254         while (*p) {
255                 int overlap = 0;
256                 parent = *p;
257
258                 __ipo = rb_entry(parent, struct io_piece, rb_node);
259                 if (ipo->file < __ipo->file)
260                         p = &(*p)->rb_left;
261                 else if (ipo->file > __ipo->file)
262                         p = &(*p)->rb_right;
263                 else if (ipo->offset < __ipo->offset) {
264                         p = &(*p)->rb_left;
265                         overlap = ipo->offset + ipo->len > __ipo->offset;
266                 }
267                 else if (ipo->offset > __ipo->offset) {
268                         p = &(*p)->rb_right;
269                         overlap = __ipo->offset + __ipo->len > ipo->offset;
270                 }
271                 else
272                         overlap = 1;
273
274                 if (overlap) {
275                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu",
276                                 __ipo->offset, __ipo->len,
277                                 ipo->offset, ipo->len);
278                         td->io_hist_len--;
279                         rb_erase(parent, &td->io_hist_tree);
280                         remove_trim_entry(td, __ipo);
281                         free(__ipo);
282                         goto restart;
283                 }
284         }
285
286         rb_link_node(&ipo->rb_node, parent, p);
287         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
288         ipo->flags |= IP_F_ONRB;
289         td->io_hist_len++;
290 }
291
292 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
293 {
294         struct io_piece *ipo = io_u->ipo;
295
296         if (td->ts.nr_block_infos) {
297                 uint32_t *info = io_u_block_info(td, io_u);
298                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
299                         if (io_u->ddir == DDIR_TRIM)
300                                 *info = BLOCK_INFO_SET_STATE(*info,
301                                                 BLOCK_STATE_TRIM_FAILURE);
302                         else if (io_u->ddir == DDIR_WRITE)
303                                 *info = BLOCK_INFO_SET_STATE(*info,
304                                                 BLOCK_STATE_WRITE_FAILURE);
305                 }
306         }
307
308         if (!ipo)
309                 return;
310
311         if (ipo->flags & IP_F_ONRB)
312                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
313         else if (ipo->flags & IP_F_ONLIST)
314                 flist_del(&ipo->list);
315
316         free(ipo);
317         io_u->ipo = NULL;
318         td->io_hist_len--;
319 }
320
321 void trim_io_piece(struct thread_data *td, const struct io_u *io_u)
322 {
323         struct io_piece *ipo = io_u->ipo;
324
325         if (!ipo)
326                 return;
327
328         ipo->len = io_u->xfer_buflen - io_u->resid;
329 }
330
331 void write_iolog_close(struct thread_data *td)
332 {
333         fflush(td->iolog_f);
334         fclose(td->iolog_f);
335         free(td->iolog_buf);
336         td->iolog_f = NULL;
337         td->iolog_buf = NULL;
338 }
339
340 /*
341  * Read version 2 iolog data. It is enhanced to include per-file logging,
342  * syncs, etc.
343  */
344 static int read_iolog2(struct thread_data *td, FILE *f)
345 {
346         unsigned long long offset;
347         unsigned int bytes;
348         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
349         char *fname, *act;
350         char *str, *p;
351         enum fio_ddir rw;
352
353         free_release_files(td);
354
355         /*
356          * Read in the read iolog and store it, reuse the infrastructure
357          * for doing verifications.
358          */
359         str = malloc(4096);
360         fname = malloc(256+16);
361         act = malloc(256+16);
362
363         reads = writes = waits = 0;
364         while ((p = fgets(str, 4096, f)) != NULL) {
365                 struct io_piece *ipo;
366                 int r;
367
368                 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
369                                                                         &bytes);
370                 if (r == 4) {
371                         /*
372                          * Check action first
373                          */
374                         if (!strcmp(act, "wait"))
375                                 rw = DDIR_WAIT;
376                         else if (!strcmp(act, "read"))
377                                 rw = DDIR_READ;
378                         else if (!strcmp(act, "write"))
379                                 rw = DDIR_WRITE;
380                         else if (!strcmp(act, "sync"))
381                                 rw = DDIR_SYNC;
382                         else if (!strcmp(act, "datasync"))
383                                 rw = DDIR_DATASYNC;
384                         else if (!strcmp(act, "trim"))
385                                 rw = DDIR_TRIM;
386                         else {
387                                 log_err("fio: bad iolog file action: %s\n",
388                                                                         act);
389                                 continue;
390                         }
391                         fileno = get_fileno(td, fname);
392                 } else if (r == 2) {
393                         rw = DDIR_INVAL;
394                         if (!strcmp(act, "add")) {
395                                 fileno = add_file(td, fname, 0, 1);
396                                 file_action = FIO_LOG_ADD_FILE;
397                                 continue;
398                         } else if (!strcmp(act, "open")) {
399                                 fileno = get_fileno(td, fname);
400                                 file_action = FIO_LOG_OPEN_FILE;
401                         } else if (!strcmp(act, "close")) {
402                                 fileno = get_fileno(td, fname);
403                                 file_action = FIO_LOG_CLOSE_FILE;
404                         } else {
405                                 log_err("fio: bad iolog file action: %s\n",
406                                                                         act);
407                                 continue;
408                         }
409                 } else {
410                         log_err("bad iolog2: %s", p);
411                         continue;
412                 }
413
414                 if (rw == DDIR_READ)
415                         reads++;
416                 else if (rw == DDIR_WRITE) {
417                         /*
418                          * Don't add a write for ro mode
419                          */
420                         if (read_only)
421                                 continue;
422                         writes++;
423                 } else if (rw == DDIR_WAIT) {
424                         waits++;
425                 } else if (rw == DDIR_INVAL) {
426                 } else if (!ddir_sync(rw)) {
427                         log_err("bad ddir: %d\n", rw);
428                         continue;
429                 }
430
431                 /*
432                  * Make note of file
433                  */
434                 ipo = malloc(sizeof(*ipo));
435                 init_ipo(ipo);
436                 ipo->ddir = rw;
437                 if (rw == DDIR_WAIT) {
438                         ipo->delay = offset;
439                 } else {
440                         ipo->offset = offset;
441                         ipo->len = bytes;
442                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw])
443                                 td->o.max_bs[rw] = bytes;
444                         ipo->fileno = fileno;
445                         ipo->file_action = file_action;
446                         td->o.size += bytes;
447                 }
448
449                 queue_io_piece(td, ipo);
450         }
451
452         free(str);
453         free(act);
454         free(fname);
455
456         if (writes && read_only) {
457                 log_err("fio: <%s> skips replay of %d writes due to"
458                         " read-only\n", td->o.name, writes);
459                 writes = 0;
460         }
461
462         if (!reads && !writes && !waits)
463                 return 1;
464         else if (reads && !writes)
465                 td->o.td_ddir = TD_DDIR_READ;
466         else if (!reads && writes)
467                 td->o.td_ddir = TD_DDIR_WRITE;
468         else
469                 td->o.td_ddir = TD_DDIR_RW;
470
471         return 0;
472 }
473
474 /*
475  * open iolog, check version, and call appropriate parser
476  */
477 static int init_iolog_read(struct thread_data *td)
478 {
479         char buffer[256], *p;
480         FILE *f;
481         int ret;
482
483         f = fopen(td->o.read_iolog_file, "r");
484         if (!f) {
485                 perror("fopen read iolog");
486                 return 1;
487         }
488
489         p = fgets(buffer, sizeof(buffer), f);
490         if (!p) {
491                 td_verror(td, errno, "iolog read");
492                 log_err("fio: unable to read iolog\n");
493                 fclose(f);
494                 return 1;
495         }
496
497         /*
498          * version 2 of the iolog stores a specific string as the
499          * first line, check for that
500          */
501         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
502                 ret = read_iolog2(td, f);
503         else {
504                 log_err("fio: iolog version 1 is no longer supported\n");
505                 ret = 1;
506         }
507
508         fclose(f);
509         return ret;
510 }
511
512 /*
513  * Set up a log for storing io patterns.
514  */
515 static int init_iolog_write(struct thread_data *td)
516 {
517         struct fio_file *ff;
518         FILE *f;
519         unsigned int i;
520
521         f = fopen(td->o.write_iolog_file, "a");
522         if (!f) {
523                 perror("fopen write iolog");
524                 return 1;
525         }
526
527         /*
528          * That's it for writing, setup a log buffer and we're done.
529           */
530         td->iolog_f = f;
531         td->iolog_buf = malloc(8192);
532         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
533
534         /*
535          * write our version line
536          */
537         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
538                 perror("iolog init\n");
539                 return 1;
540         }
541
542         /*
543          * add all known files
544          */
545         for_each_file(td, ff, i)
546                 log_file(td, ff, FIO_LOG_ADD_FILE);
547
548         return 0;
549 }
550
551 int init_iolog(struct thread_data *td)
552 {
553         int ret = 0;
554
555         if (td->o.read_iolog_file) {
556                 int need_swap;
557
558                 /*
559                  * Check if it's a blktrace file and load that if possible.
560                  * Otherwise assume it's a normal log file and load that.
561                  */
562                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
563                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
564                 else
565                         ret = init_iolog_read(td);
566         } else if (td->o.write_iolog_file)
567                 ret = init_iolog_write(td);
568
569         if (ret)
570                 td_verror(td, EINVAL, "failed initializing iolog");
571
572         return ret;
573 }
574
575 void setup_log(struct io_log **log, struct log_params *p,
576                const char *filename)
577 {
578         struct io_log *l;
579         int i;
580         struct io_u_plat_entry *entry;
581         struct flist_head *list;
582
583         l = scalloc(1, sizeof(*l));
584         INIT_FLIST_HEAD(&l->io_logs);
585         l->log_type = p->log_type;
586         l->log_offset = p->log_offset;
587         l->log_gz = p->log_gz;
588         l->log_gz_store = p->log_gz_store;
589         l->avg_msec = p->avg_msec;
590         l->hist_msec = p->hist_msec;
591         l->hist_coarseness = p->hist_coarseness;
592         l->filename = strdup(filename);
593         l->td = p->td;
594
595         /* Initialize histogram lists for each r/w direction,
596          * with initial io_u_plat of all zeros:
597          */
598         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
599                 list = &l->hist_window[i].list;
600                 INIT_FLIST_HEAD(list);
601                 entry = calloc(1, sizeof(struct io_u_plat_entry));
602                 flist_add(&entry->list, list);
603         }
604
605         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
606                 struct io_logs *p;
607
608                 p = calloc(1, sizeof(*l->pending));
609                 p->max_samples = DEF_LOG_ENTRIES;
610                 p->log = calloc(p->max_samples, log_entry_sz(l));
611                 l->pending = p;
612         }
613
614         if (l->log_offset)
615                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
616
617         INIT_FLIST_HEAD(&l->chunk_list);
618
619         if (l->log_gz && !p->td)
620                 l->log_gz = 0;
621         else if (l->log_gz || l->log_gz_store) {
622                 mutex_init_pshared(&l->chunk_lock);
623                 p->td->flags |= TD_F_COMPRESS_LOG;
624         }
625
626         *log = l;
627 }
628
629 #ifdef CONFIG_SETVBUF
630 static void *set_file_buffer(FILE *f)
631 {
632         size_t size = 1048576;
633         void *buf;
634
635         buf = malloc(size);
636         setvbuf(f, buf, _IOFBF, size);
637         return buf;
638 }
639
640 static void clear_file_buffer(void *buf)
641 {
642         free(buf);
643 }
644 #else
645 static void *set_file_buffer(FILE *f)
646 {
647         return NULL;
648 }
649
650 static void clear_file_buffer(void *buf)
651 {
652 }
653 #endif
654
655 void free_log(struct io_log *log)
656 {
657         while (!flist_empty(&log->io_logs)) {
658                 struct io_logs *cur_log;
659
660                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
661                 flist_del_init(&cur_log->list);
662                 free(cur_log->log);
663                 sfree(cur_log);
664         }
665
666         if (log->pending) {
667                 free(log->pending->log);
668                 free(log->pending);
669                 log->pending = NULL;
670         }
671
672         free(log->pending);
673         free(log->filename);
674         sfree(log);
675 }
676
677 static inline unsigned long hist_sum(int j, int stride, unsigned int *io_u_plat,
678                 unsigned int *io_u_plat_last)
679 {
680         unsigned long sum;
681         int k;
682
683         for (k = sum = 0; k < stride; k++)
684                 sum += io_u_plat[j + k] - io_u_plat_last[j + k];
685
686         return sum;
687 }
688
689 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
690                                uint64_t sample_size)
691 {
692         struct io_sample *s;
693         int log_offset;
694         uint64_t i, j, nr_samples;
695         struct io_u_plat_entry *entry, *entry_before;
696         unsigned int *io_u_plat;
697         unsigned int *io_u_plat_before;
698
699         int stride = 1 << hist_coarseness;
700         
701         if (!sample_size)
702                 return;
703
704         s = __get_sample(samples, 0, 0);
705         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
706
707         nr_samples = sample_size / __log_entry_sz(log_offset);
708
709         for (i = 0; i < nr_samples; i++) {
710                 s = __get_sample(samples, log_offset, i);
711
712                 entry = (struct io_u_plat_entry *) s->val;
713                 io_u_plat = entry->io_u_plat;
714
715                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
716                 io_u_plat_before = entry_before->io_u_plat;
717
718                 fprintf(f, "%lu, %u, %u, ", (unsigned long) s->time,
719                                                 io_sample_ddir(s), s->bs);
720                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
721                         fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat,
722                                                 io_u_plat_before));
723                 }
724                 fprintf(f, "%lu\n", (unsigned long)
725                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
726                                         io_u_plat_before));
727
728                 flist_del(&entry_before->list);
729                 free(entry_before);
730         }
731 }
732
733 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
734 {
735         struct io_sample *s;
736         int log_offset;
737         uint64_t i, nr_samples;
738
739         if (!sample_size)
740                 return;
741
742         s = __get_sample(samples, 0, 0);
743         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
744
745         nr_samples = sample_size / __log_entry_sz(log_offset);
746
747         for (i = 0; i < nr_samples; i++) {
748                 s = __get_sample(samples, log_offset, i);
749
750                 if (!log_offset) {
751                         fprintf(f, "%lu, %lu, %u, %u\n",
752                                         (unsigned long) s->time,
753                                         (unsigned long) s->val,
754                                         io_sample_ddir(s), s->bs);
755                 } else {
756                         struct io_sample_offset *so = (void *) s;
757
758                         fprintf(f, "%lu, %lu, %u, %u, %llu\n",
759                                         (unsigned long) s->time,
760                                         (unsigned long) s->val,
761                                         io_sample_ddir(s), s->bs,
762                                         (unsigned long long) so->offset);
763                 }
764         }
765 }
766
767 #ifdef CONFIG_ZLIB
768
769 struct iolog_flush_data {
770         struct workqueue_work work;
771         struct io_log *log;
772         void *samples;
773         uint32_t nr_samples;
774         bool free;
775 };
776
777 #define GZ_CHUNK        131072
778
779 static struct iolog_compress *get_new_chunk(unsigned int seq)
780 {
781         struct iolog_compress *c;
782
783         c = malloc(sizeof(*c));
784         INIT_FLIST_HEAD(&c->list);
785         c->buf = malloc(GZ_CHUNK);
786         c->len = 0;
787         c->seq = seq;
788         return c;
789 }
790
791 static void free_chunk(struct iolog_compress *ic)
792 {
793         free(ic->buf);
794         free(ic);
795 }
796
797 static int z_stream_init(z_stream *stream, int gz_hdr)
798 {
799         int wbits = 15;
800
801         memset(stream, 0, sizeof(*stream));
802         stream->zalloc = Z_NULL;
803         stream->zfree = Z_NULL;
804         stream->opaque = Z_NULL;
805         stream->next_in = Z_NULL;
806
807         /*
808          * zlib magic - add 32 for auto-detection of gz header or not,
809          * if we decide to store files in a gzip friendly format.
810          */
811         if (gz_hdr)
812                 wbits += 32;
813
814         if (inflateInit2(stream, wbits) != Z_OK)
815                 return 1;
816
817         return 0;
818 }
819
820 struct inflate_chunk_iter {
821         unsigned int seq;
822         int err;
823         void *buf;
824         size_t buf_size;
825         size_t buf_used;
826         size_t chunk_sz;
827 };
828
829 static void finish_chunk(z_stream *stream, FILE *f,
830                          struct inflate_chunk_iter *iter)
831 {
832         int ret;
833
834         ret = inflateEnd(stream);
835         if (ret != Z_OK)
836                 log_err("fio: failed to end log inflation seq %d (%d)\n",
837                                 iter->seq, ret);
838
839         flush_samples(f, iter->buf, iter->buf_used);
840         free(iter->buf);
841         iter->buf = NULL;
842         iter->buf_size = iter->buf_used = 0;
843 }
844
845 /*
846  * Iterative chunk inflation. Handles cases where we cross into a new
847  * sequence, doing flush finish of previous chunk if needed.
848  */
849 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
850                             z_stream *stream, struct inflate_chunk_iter *iter)
851 {
852         size_t ret;
853
854         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
855                                 (unsigned long) ic->len, ic->seq);
856
857         if (ic->seq != iter->seq) {
858                 if (iter->seq)
859                         finish_chunk(stream, f, iter);
860
861                 z_stream_init(stream, gz_hdr);
862                 iter->seq = ic->seq;
863         }
864
865         stream->avail_in = ic->len;
866         stream->next_in = ic->buf;
867
868         if (!iter->buf_size) {
869                 iter->buf_size = iter->chunk_sz;
870                 iter->buf = malloc(iter->buf_size);
871         }
872
873         while (stream->avail_in) {
874                 size_t this_out = iter->buf_size - iter->buf_used;
875                 int err;
876
877                 stream->avail_out = this_out;
878                 stream->next_out = iter->buf + iter->buf_used;
879
880                 err = inflate(stream, Z_NO_FLUSH);
881                 if (err < 0) {
882                         log_err("fio: failed inflating log: %d\n", err);
883                         iter->err = err;
884                         break;
885                 }
886
887                 iter->buf_used += this_out - stream->avail_out;
888
889                 if (!stream->avail_out) {
890                         iter->buf_size += iter->chunk_sz;
891                         iter->buf = realloc(iter->buf, iter->buf_size);
892                         continue;
893                 }
894
895                 if (err == Z_STREAM_END)
896                         break;
897         }
898
899         ret = (void *) stream->next_in - ic->buf;
900
901         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
902
903         return ret;
904 }
905
906 /*
907  * Inflate stored compressed chunks, or write them directly to the log
908  * file if so instructed.
909  */
910 static int inflate_gz_chunks(struct io_log *log, FILE *f)
911 {
912         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
913         z_stream stream;
914
915         while (!flist_empty(&log->chunk_list)) {
916                 struct iolog_compress *ic;
917
918                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
919                 flist_del(&ic->list);
920
921                 if (log->log_gz_store) {
922                         size_t ret;
923
924                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
925                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
926
927                         ret = fwrite(ic->buf, ic->len, 1, f);
928                         if (ret != 1 || ferror(f)) {
929                                 iter.err = errno;
930                                 log_err("fio: error writing compressed log\n");
931                         }
932                 } else
933                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
934
935                 free_chunk(ic);
936         }
937
938         if (iter.seq) {
939                 finish_chunk(&stream, f, &iter);
940                 free(iter.buf);
941         }
942
943         return iter.err;
944 }
945
946 /*
947  * Open compressed log file and decompress the stored chunks and
948  * write them to stdout. The chunks are stored sequentially in the
949  * file, so we iterate over them and do them one-by-one.
950  */
951 int iolog_file_inflate(const char *file)
952 {
953         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
954         struct iolog_compress ic;
955         z_stream stream;
956         struct stat sb;
957         ssize_t ret;
958         size_t total;
959         void *buf;
960         FILE *f;
961
962         f = fopen(file, "r");
963         if (!f) {
964                 perror("fopen");
965                 return 1;
966         }
967
968         if (stat(file, &sb) < 0) {
969                 fclose(f);
970                 perror("stat");
971                 return 1;
972         }
973
974         ic.buf = buf = malloc(sb.st_size);
975         ic.len = sb.st_size;
976         ic.seq = 1;
977
978         ret = fread(ic.buf, ic.len, 1, f);
979         if (ret < 0) {
980                 perror("fread");
981                 fclose(f);
982                 free(buf);
983                 return 1;
984         } else if (ret != 1) {
985                 log_err("fio: short read on reading log\n");
986                 fclose(f);
987                 free(buf);
988                 return 1;
989         }
990
991         fclose(f);
992
993         /*
994          * Each chunk will return Z_STREAM_END. We don't know how many
995          * chunks are in the file, so we just keep looping and incrementing
996          * the sequence number until we have consumed the whole compressed
997          * file.
998          */
999         total = ic.len;
1000         do {
1001                 size_t iret;
1002
1003                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1004                 total -= iret;
1005                 if (!total)
1006                         break;
1007                 if (iter.err)
1008                         break;
1009
1010                 ic.seq++;
1011                 ic.len -= iret;
1012                 ic.buf += iret;
1013         } while (1);
1014
1015         if (iter.seq) {
1016                 finish_chunk(&stream, stdout, &iter);
1017                 free(iter.buf);
1018         }
1019
1020         free(buf);
1021         return iter.err;
1022 }
1023
1024 #else
1025
1026 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1027 {
1028         return 0;
1029 }
1030
1031 int iolog_file_inflate(const char *file)
1032 {
1033         log_err("fio: log inflation not possible without zlib\n");
1034         return 1;
1035 }
1036
1037 #endif
1038
1039 void flush_log(struct io_log *log, bool do_append)
1040 {
1041         void *buf;
1042         FILE *f;
1043
1044         if (!do_append)
1045                 f = fopen(log->filename, "w");
1046         else
1047                 f = fopen(log->filename, "a");
1048         if (!f) {
1049                 perror("fopen log");
1050                 return;
1051         }
1052
1053         buf = set_file_buffer(f);
1054
1055         inflate_gz_chunks(log, f);
1056
1057         while (!flist_empty(&log->io_logs)) {
1058                 struct io_logs *cur_log;
1059
1060                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1061                 flist_del_init(&cur_log->list);
1062                 
1063                 if (log == log->td->clat_hist_log)
1064                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1065                                            cur_log->nr_samples * log_entry_sz(log));
1066                 else
1067                         flush_samples(f, cur_log->log, cur_log->nr_samples * log_entry_sz(log));
1068                 
1069                 sfree(cur_log);
1070         }
1071
1072         fclose(f);
1073         clear_file_buffer(buf);
1074 }
1075
1076 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1077 {
1078         if (td->flags & TD_F_COMPRESS_LOG)
1079                 iolog_flush(log);
1080
1081         if (trylock) {
1082                 if (fio_trylock_file(log->filename))
1083                         return 1;
1084         } else
1085                 fio_lock_file(log->filename);
1086
1087         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1088                 fio_send_iolog(td, log, log->filename);
1089         else
1090                 flush_log(log, !td->o.per_job_logs);
1091
1092         fio_unlock_file(log->filename);
1093         free_log(log);
1094         return 0;
1095 }
1096
1097 size_t log_chunk_sizes(struct io_log *log)
1098 {
1099         struct flist_head *entry;
1100         size_t ret;
1101
1102         if (flist_empty(&log->chunk_list))
1103                 return 0;
1104
1105         ret = 0;
1106         pthread_mutex_lock(&log->chunk_lock);
1107         flist_for_each(entry, &log->chunk_list) {
1108                 struct iolog_compress *c;
1109
1110                 c = flist_entry(entry, struct iolog_compress, list);
1111                 ret += c->len;
1112         }
1113         pthread_mutex_unlock(&log->chunk_lock);
1114         return ret;
1115 }
1116
1117 #ifdef CONFIG_ZLIB
1118
1119 static int gz_work(struct iolog_flush_data *data)
1120 {
1121         struct iolog_compress *c = NULL;
1122         struct flist_head list;
1123         unsigned int seq;
1124         z_stream stream;
1125         size_t total = 0;
1126         int ret;
1127
1128         INIT_FLIST_HEAD(&list);
1129
1130         memset(&stream, 0, sizeof(stream));
1131         stream.zalloc = Z_NULL;
1132         stream.zfree = Z_NULL;
1133         stream.opaque = Z_NULL;
1134
1135         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1136         if (ret != Z_OK) {
1137                 log_err("fio: failed to init gz stream\n");
1138                 goto err;
1139         }
1140
1141         seq = ++data->log->chunk_seq;
1142
1143         stream.next_in = (void *) data->samples;
1144         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1145
1146         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1147                                 (unsigned long) stream.avail_in, seq,
1148                                 data->log->filename);
1149         do {
1150                 if (c)
1151                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1152                 c = get_new_chunk(seq);
1153                 stream.avail_out = GZ_CHUNK;
1154                 stream.next_out = c->buf;
1155                 ret = deflate(&stream, Z_NO_FLUSH);
1156                 if (ret < 0) {
1157                         log_err("fio: deflate log (%d)\n", ret);
1158                         free_chunk(c);
1159                         goto err;
1160                 }
1161
1162                 c->len = GZ_CHUNK - stream.avail_out;
1163                 flist_add_tail(&c->list, &list);
1164                 total += c->len;
1165         } while (stream.avail_in);
1166
1167         stream.next_out = c->buf + c->len;
1168         stream.avail_out = GZ_CHUNK - c->len;
1169
1170         ret = deflate(&stream, Z_FINISH);
1171         if (ret < 0) {
1172                 /*
1173                  * Z_BUF_ERROR is special, it just means we need more
1174                  * output space. We'll handle that below. Treat any other
1175                  * error as fatal.
1176                  */
1177                 if (ret != Z_BUF_ERROR) {
1178                         log_err("fio: deflate log (%d)\n", ret);
1179                         flist_del(&c->list);
1180                         free_chunk(c);
1181                         goto err;
1182                 }
1183         }
1184
1185         total -= c->len;
1186         c->len = GZ_CHUNK - stream.avail_out;
1187         total += c->len;
1188         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1189
1190         if (ret != Z_STREAM_END) {
1191                 do {
1192                         c = get_new_chunk(seq);
1193                         stream.avail_out = GZ_CHUNK;
1194                         stream.next_out = c->buf;
1195                         ret = deflate(&stream, Z_FINISH);
1196                         c->len = GZ_CHUNK - stream.avail_out;
1197                         total += c->len;
1198                         flist_add_tail(&c->list, &list);
1199                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1200                 } while (ret != Z_STREAM_END);
1201         }
1202
1203         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1204
1205         ret = deflateEnd(&stream);
1206         if (ret != Z_OK)
1207                 log_err("fio: deflateEnd %d\n", ret);
1208
1209         free(data->samples);
1210
1211         if (!flist_empty(&list)) {
1212                 pthread_mutex_lock(&data->log->chunk_lock);
1213                 flist_splice_tail(&list, &data->log->chunk_list);
1214                 pthread_mutex_unlock(&data->log->chunk_lock);
1215         }
1216
1217         ret = 0;
1218 done:
1219         if (data->free)
1220                 free(data);
1221         return ret;
1222 err:
1223         while (!flist_empty(&list)) {
1224                 c = flist_first_entry(list.next, struct iolog_compress, list);
1225                 flist_del(&c->list);
1226                 free_chunk(c);
1227         }
1228         ret = 1;
1229         goto done;
1230 }
1231
1232 /*
1233  * Invoked from our compress helper thread, when logging would have exceeded
1234  * the specified memory limitation. Compresses the previously stored
1235  * entries.
1236  */
1237 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1238 {
1239         return gz_work(container_of(work, struct iolog_flush_data, work));
1240 }
1241
1242 static int gz_init_worker(struct submit_worker *sw)
1243 {
1244         struct thread_data *td = sw->wq->td;
1245
1246         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1247                 return 0;
1248
1249         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1250                 log_err("gz: failed to set CPU affinity\n");
1251                 return 1;
1252         }
1253
1254         return 0;
1255 }
1256
1257 static struct workqueue_ops log_compress_wq_ops = {
1258         .fn             = gz_work_async,
1259         .init_worker_fn = gz_init_worker,
1260         .nice           = 1,
1261 };
1262
1263 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1264 {
1265         if (!(td->flags & TD_F_COMPRESS_LOG))
1266                 return 0;
1267
1268         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1269         return 0;
1270 }
1271
1272 void iolog_compress_exit(struct thread_data *td)
1273 {
1274         if (!(td->flags & TD_F_COMPRESS_LOG))
1275                 return;
1276
1277         workqueue_exit(&td->log_compress_wq);
1278 }
1279
1280 /*
1281  * Queue work item to compress the existing log entries. We reset the
1282  * current log to a small size, and reference the existing log in the
1283  * data that we queue for compression. Once compression has been done,
1284  * this old log is freed. If called with finish == true, will not return
1285  * until the log compression has completed, and will flush all previous
1286  * logs too
1287  */
1288 static int iolog_flush(struct io_log *log)
1289 {
1290         struct iolog_flush_data *data;
1291
1292         data = malloc(sizeof(*data));
1293         if (!data)
1294                 return 1;
1295
1296         data->log = log;
1297         data->free = false;
1298
1299         while (!flist_empty(&log->io_logs)) {
1300                 struct io_logs *cur_log;
1301
1302                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1303                 flist_del_init(&cur_log->list);
1304
1305                 data->samples = cur_log->log;
1306                 data->nr_samples = cur_log->nr_samples;
1307
1308                 sfree(cur_log);
1309
1310                 gz_work(data);
1311         }
1312
1313         free(data);
1314         return 0;
1315 }
1316
1317 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1318 {
1319         struct iolog_flush_data *data;
1320
1321         data = malloc(sizeof(*data));
1322         if (!data)
1323                 return 1;
1324
1325         data->log = log;
1326
1327         data->samples = cur_log->log;
1328         data->nr_samples = cur_log->nr_samples;
1329         data->free = true;
1330
1331         cur_log->nr_samples = cur_log->max_samples = 0;
1332         cur_log->log = NULL;
1333
1334         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1335         return 0;
1336 }
1337 #else
1338
1339 static int iolog_flush(struct io_log *log)
1340 {
1341         return 1;
1342 }
1343
1344 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1345 {
1346         return 1;
1347 }
1348
1349 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1350 {
1351         return 0;
1352 }
1353
1354 void iolog_compress_exit(struct thread_data *td)
1355 {
1356 }
1357
1358 #endif
1359
1360 struct io_logs *iolog_cur_log(struct io_log *log)
1361 {
1362         if (flist_empty(&log->io_logs))
1363                 return NULL;
1364
1365         return flist_last_entry(&log->io_logs, struct io_logs, list);
1366 }
1367
1368 uint64_t iolog_nr_samples(struct io_log *iolog)
1369 {
1370         struct flist_head *entry;
1371         uint64_t ret = 0;
1372
1373         flist_for_each(entry, &iolog->io_logs) {
1374                 struct io_logs *cur_log;
1375
1376                 cur_log = flist_entry(entry, struct io_logs, list);
1377                 ret += cur_log->nr_samples;
1378         }
1379
1380         return ret;
1381 }
1382
1383 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1384 {
1385         if (log)
1386                 return finish_log(td, log, try);
1387
1388         return 0;
1389 }
1390
1391 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1392 {
1393         int ret;
1394
1395         if (per_unit_log(td->iops_log) != unit_log)
1396                 return 0;
1397
1398         ret = __write_log(td, td->iops_log, try);
1399         if (!ret)
1400                 td->iops_log = NULL;
1401
1402         return ret;
1403 }
1404
1405 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1406 {
1407         int ret;
1408
1409         if (!unit_log)
1410                 return 0;
1411
1412         ret = __write_log(td, td->slat_log, try);
1413         if (!ret)
1414                 td->slat_log = NULL;
1415
1416         return ret;
1417 }
1418
1419 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1420 {
1421         int ret;
1422
1423         if (!unit_log)
1424                 return 0;
1425
1426         ret = __write_log(td, td->clat_log, try);
1427         if (!ret)
1428                 td->clat_log = NULL;
1429
1430         return ret;
1431 }
1432
1433 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1434 {
1435         int ret;
1436
1437         if (!unit_log)
1438                 return 0;
1439
1440         ret = __write_log(td, td->clat_hist_log, try);
1441         if (!ret)
1442                 td->clat_hist_log = NULL;
1443
1444         return ret;
1445 }
1446
1447 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1448 {
1449         int ret;
1450
1451         if (!unit_log)
1452                 return 0;
1453
1454         ret = __write_log(td, td->lat_log, try);
1455         if (!ret)
1456                 td->lat_log = NULL;
1457
1458         return ret;
1459 }
1460
1461 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1462 {
1463         int ret;
1464
1465         if (per_unit_log(td->bw_log) != unit_log)
1466                 return 0;
1467
1468         ret = __write_log(td, td->bw_log, try);
1469         if (!ret)
1470                 td->bw_log = NULL;
1471
1472         return ret;
1473 }
1474
1475 enum {
1476         BW_LOG_MASK     = 1,
1477         LAT_LOG_MASK    = 2,
1478         SLAT_LOG_MASK   = 4,
1479         CLAT_LOG_MASK   = 8,
1480         IOPS_LOG_MASK   = 16,
1481         CLAT_HIST_LOG_MASK = 32,
1482
1483         ALL_LOG_NR      = 6,
1484 };
1485
1486 struct log_type {
1487         unsigned int mask;
1488         int (*fn)(struct thread_data *, int, bool);
1489 };
1490
1491 static struct log_type log_types[] = {
1492         {
1493                 .mask   = BW_LOG_MASK,
1494                 .fn     = write_bandw_log,
1495         },
1496         {
1497                 .mask   = LAT_LOG_MASK,
1498                 .fn     = write_lat_log,
1499         },
1500         {
1501                 .mask   = SLAT_LOG_MASK,
1502                 .fn     = write_slat_log,
1503         },
1504         {
1505                 .mask   = CLAT_LOG_MASK,
1506                 .fn     = write_clat_log,
1507         },
1508         {
1509                 .mask   = IOPS_LOG_MASK,
1510                 .fn     = write_iops_log,
1511         },
1512         {
1513                 .mask   = CLAT_HIST_LOG_MASK,
1514                 .fn     = write_clat_hist_log,
1515         }
1516 };
1517
1518 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1519 {
1520         unsigned int log_mask = 0;
1521         unsigned int log_left = ALL_LOG_NR;
1522         int old_state, i;
1523
1524         old_state = td_bump_runstate(td, TD_FINISHING);
1525
1526         finalize_logs(td, unit_logs);
1527
1528         while (log_left) {
1529                 int prev_log_left = log_left;
1530
1531                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1532                         struct log_type *lt = &log_types[i];
1533                         int ret;
1534
1535                         if (!(log_mask & lt->mask)) {
1536                                 ret = lt->fn(td, log_left != 1, unit_logs);
1537                                 if (!ret) {
1538                                         log_left--;
1539                                         log_mask |= lt->mask;
1540                                 }
1541                         }
1542                 }
1543
1544                 if (prev_log_left == log_left)
1545                         usleep(5000);
1546         }
1547
1548         td_restore_runstate(td, old_state);
1549 }
1550
1551 void fio_writeout_logs(bool unit_logs)
1552 {
1553         struct thread_data *td;
1554         int i;
1555
1556         for_each_td(td, i)
1557                 td_writeout_logs(td, unit_logs);
1558 }