bloom: add string version
[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 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         if (io_u_plat_last) {
684                 for (k = sum = 0; k < stride; k++)
685                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
686         } else {
687                 for (k = sum = 0; k < stride; k++)
688                         sum += io_u_plat[j + k];
689         }
690
691         return sum;
692 }
693
694 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
695                                uint64_t sample_size)
696 {
697         struct io_sample *s;
698         int log_offset;
699         uint64_t i, j, nr_samples;
700         struct io_u_plat_entry *entry, *entry_before;
701         unsigned int *io_u_plat;
702         unsigned int *io_u_plat_before;
703
704         int stride = 1 << hist_coarseness;
705         
706         if (!sample_size)
707                 return;
708
709         s = __get_sample(samples, 0, 0);
710         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
711
712         nr_samples = sample_size / __log_entry_sz(log_offset);
713
714         for (i = 0; i < nr_samples; i++) {
715                 s = __get_sample(samples, log_offset, i);
716
717                 entry = (struct io_u_plat_entry *) (uintptr_t) s->val;
718                 io_u_plat = entry->io_u_plat;
719
720                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
721                 io_u_plat_before = entry_before->io_u_plat;
722
723                 fprintf(f, "%lu, %u, %u, ", (unsigned long) s->time,
724                                                 io_sample_ddir(s), s->bs);
725                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
726                         fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat,
727                                                 io_u_plat_before));
728                 }
729                 fprintf(f, "%lu\n", (unsigned long)
730                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
731                                         io_u_plat_before));
732
733                 flist_del(&entry_before->list);
734                 free(entry_before);
735         }
736 }
737
738 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
739 {
740         struct io_sample *s;
741         int log_offset;
742         uint64_t i, nr_samples;
743
744         if (!sample_size)
745                 return;
746
747         s = __get_sample(samples, 0, 0);
748         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
749
750         nr_samples = sample_size / __log_entry_sz(log_offset);
751
752         for (i = 0; i < nr_samples; i++) {
753                 s = __get_sample(samples, log_offset, i);
754
755                 if (!log_offset) {
756                         fprintf(f, "%lu, %lu, %u, %u\n",
757                                         (unsigned long) s->time,
758                                         (unsigned long) s->val,
759                                         io_sample_ddir(s), s->bs);
760                 } else {
761                         struct io_sample_offset *so = (void *) s;
762
763                         fprintf(f, "%lu, %lu, %u, %u, %llu\n",
764                                         (unsigned long) s->time,
765                                         (unsigned long) s->val,
766                                         io_sample_ddir(s), s->bs,
767                                         (unsigned long long) so->offset);
768                 }
769         }
770 }
771
772 #ifdef CONFIG_ZLIB
773
774 struct iolog_flush_data {
775         struct workqueue_work work;
776         struct io_log *log;
777         void *samples;
778         uint32_t nr_samples;
779         bool free;
780 };
781
782 #define GZ_CHUNK        131072
783
784 static struct iolog_compress *get_new_chunk(unsigned int seq)
785 {
786         struct iolog_compress *c;
787
788         c = malloc(sizeof(*c));
789         INIT_FLIST_HEAD(&c->list);
790         c->buf = malloc(GZ_CHUNK);
791         c->len = 0;
792         c->seq = seq;
793         return c;
794 }
795
796 static void free_chunk(struct iolog_compress *ic)
797 {
798         free(ic->buf);
799         free(ic);
800 }
801
802 static int z_stream_init(z_stream *stream, int gz_hdr)
803 {
804         int wbits = 15;
805
806         memset(stream, 0, sizeof(*stream));
807         stream->zalloc = Z_NULL;
808         stream->zfree = Z_NULL;
809         stream->opaque = Z_NULL;
810         stream->next_in = Z_NULL;
811
812         /*
813          * zlib magic - add 32 for auto-detection of gz header or not,
814          * if we decide to store files in a gzip friendly format.
815          */
816         if (gz_hdr)
817                 wbits += 32;
818
819         if (inflateInit2(stream, wbits) != Z_OK)
820                 return 1;
821
822         return 0;
823 }
824
825 struct inflate_chunk_iter {
826         unsigned int seq;
827         int err;
828         void *buf;
829         size_t buf_size;
830         size_t buf_used;
831         size_t chunk_sz;
832 };
833
834 static void finish_chunk(z_stream *stream, FILE *f,
835                          struct inflate_chunk_iter *iter)
836 {
837         int ret;
838
839         ret = inflateEnd(stream);
840         if (ret != Z_OK)
841                 log_err("fio: failed to end log inflation seq %d (%d)\n",
842                                 iter->seq, ret);
843
844         flush_samples(f, iter->buf, iter->buf_used);
845         free(iter->buf);
846         iter->buf = NULL;
847         iter->buf_size = iter->buf_used = 0;
848 }
849
850 /*
851  * Iterative chunk inflation. Handles cases where we cross into a new
852  * sequence, doing flush finish of previous chunk if needed.
853  */
854 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
855                             z_stream *stream, struct inflate_chunk_iter *iter)
856 {
857         size_t ret;
858
859         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
860                                 (unsigned long) ic->len, ic->seq);
861
862         if (ic->seq != iter->seq) {
863                 if (iter->seq)
864                         finish_chunk(stream, f, iter);
865
866                 z_stream_init(stream, gz_hdr);
867                 iter->seq = ic->seq;
868         }
869
870         stream->avail_in = ic->len;
871         stream->next_in = ic->buf;
872
873         if (!iter->buf_size) {
874                 iter->buf_size = iter->chunk_sz;
875                 iter->buf = malloc(iter->buf_size);
876         }
877
878         while (stream->avail_in) {
879                 size_t this_out = iter->buf_size - iter->buf_used;
880                 int err;
881
882                 stream->avail_out = this_out;
883                 stream->next_out = iter->buf + iter->buf_used;
884
885                 err = inflate(stream, Z_NO_FLUSH);
886                 if (err < 0) {
887                         log_err("fio: failed inflating log: %d\n", err);
888                         iter->err = err;
889                         break;
890                 }
891
892                 iter->buf_used += this_out - stream->avail_out;
893
894                 if (!stream->avail_out) {
895                         iter->buf_size += iter->chunk_sz;
896                         iter->buf = realloc(iter->buf, iter->buf_size);
897                         continue;
898                 }
899
900                 if (err == Z_STREAM_END)
901                         break;
902         }
903
904         ret = (void *) stream->next_in - ic->buf;
905
906         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
907
908         return ret;
909 }
910
911 /*
912  * Inflate stored compressed chunks, or write them directly to the log
913  * file if so instructed.
914  */
915 static int inflate_gz_chunks(struct io_log *log, FILE *f)
916 {
917         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
918         z_stream stream;
919
920         while (!flist_empty(&log->chunk_list)) {
921                 struct iolog_compress *ic;
922
923                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
924                 flist_del(&ic->list);
925
926                 if (log->log_gz_store) {
927                         size_t ret;
928
929                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
930                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
931
932                         ret = fwrite(ic->buf, ic->len, 1, f);
933                         if (ret != 1 || ferror(f)) {
934                                 iter.err = errno;
935                                 log_err("fio: error writing compressed log\n");
936                         }
937                 } else
938                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
939
940                 free_chunk(ic);
941         }
942
943         if (iter.seq) {
944                 finish_chunk(&stream, f, &iter);
945                 free(iter.buf);
946         }
947
948         return iter.err;
949 }
950
951 /*
952  * Open compressed log file and decompress the stored chunks and
953  * write them to stdout. The chunks are stored sequentially in the
954  * file, so we iterate over them and do them one-by-one.
955  */
956 int iolog_file_inflate(const char *file)
957 {
958         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
959         struct iolog_compress ic;
960         z_stream stream;
961         struct stat sb;
962         ssize_t ret;
963         size_t total;
964         void *buf;
965         FILE *f;
966
967         f = fopen(file, "r");
968         if (!f) {
969                 perror("fopen");
970                 return 1;
971         }
972
973         if (stat(file, &sb) < 0) {
974                 fclose(f);
975                 perror("stat");
976                 return 1;
977         }
978
979         ic.buf = buf = malloc(sb.st_size);
980         ic.len = sb.st_size;
981         ic.seq = 1;
982
983         ret = fread(ic.buf, ic.len, 1, f);
984         if (ret < 0) {
985                 perror("fread");
986                 fclose(f);
987                 free(buf);
988                 return 1;
989         } else if (ret != 1) {
990                 log_err("fio: short read on reading log\n");
991                 fclose(f);
992                 free(buf);
993                 return 1;
994         }
995
996         fclose(f);
997
998         /*
999          * Each chunk will return Z_STREAM_END. We don't know how many
1000          * chunks are in the file, so we just keep looping and incrementing
1001          * the sequence number until we have consumed the whole compressed
1002          * file.
1003          */
1004         total = ic.len;
1005         do {
1006                 size_t iret;
1007
1008                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1009                 total -= iret;
1010                 if (!total)
1011                         break;
1012                 if (iter.err)
1013                         break;
1014
1015                 ic.seq++;
1016                 ic.len -= iret;
1017                 ic.buf += iret;
1018         } while (1);
1019
1020         if (iter.seq) {
1021                 finish_chunk(&stream, stdout, &iter);
1022                 free(iter.buf);
1023         }
1024
1025         free(buf);
1026         return iter.err;
1027 }
1028
1029 #else
1030
1031 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1032 {
1033         return 0;
1034 }
1035
1036 int iolog_file_inflate(const char *file)
1037 {
1038         log_err("fio: log inflation not possible without zlib\n");
1039         return 1;
1040 }
1041
1042 #endif
1043
1044 void flush_log(struct io_log *log, bool do_append)
1045 {
1046         void *buf;
1047         FILE *f;
1048
1049         if (!do_append)
1050                 f = fopen(log->filename, "w");
1051         else
1052                 f = fopen(log->filename, "a");
1053         if (!f) {
1054                 perror("fopen log");
1055                 return;
1056         }
1057
1058         buf = set_file_buffer(f);
1059
1060         inflate_gz_chunks(log, f);
1061
1062         while (!flist_empty(&log->io_logs)) {
1063                 struct io_logs *cur_log;
1064
1065                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1066                 flist_del_init(&cur_log->list);
1067                 
1068                 if (log == log->td->clat_hist_log)
1069                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1070                                            log_sample_sz(log, cur_log));
1071                 else
1072                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1073                 
1074                 sfree(cur_log);
1075         }
1076
1077         fclose(f);
1078         clear_file_buffer(buf);
1079 }
1080
1081 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1082 {
1083         if (td->flags & TD_F_COMPRESS_LOG)
1084                 iolog_flush(log);
1085
1086         if (trylock) {
1087                 if (fio_trylock_file(log->filename))
1088                         return 1;
1089         } else
1090                 fio_lock_file(log->filename);
1091
1092         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1093                 fio_send_iolog(td, log, log->filename);
1094         else
1095                 flush_log(log, !td->o.per_job_logs);
1096
1097         fio_unlock_file(log->filename);
1098         free_log(log);
1099         return 0;
1100 }
1101
1102 size_t log_chunk_sizes(struct io_log *log)
1103 {
1104         struct flist_head *entry;
1105         size_t ret;
1106
1107         if (flist_empty(&log->chunk_list))
1108                 return 0;
1109
1110         ret = 0;
1111         pthread_mutex_lock(&log->chunk_lock);
1112         flist_for_each(entry, &log->chunk_list) {
1113                 struct iolog_compress *c;
1114
1115                 c = flist_entry(entry, struct iolog_compress, list);
1116                 ret += c->len;
1117         }
1118         pthread_mutex_unlock(&log->chunk_lock);
1119         return ret;
1120 }
1121
1122 #ifdef CONFIG_ZLIB
1123
1124 static int gz_work(struct iolog_flush_data *data)
1125 {
1126         struct iolog_compress *c = NULL;
1127         struct flist_head list;
1128         unsigned int seq;
1129         z_stream stream;
1130         size_t total = 0;
1131         int ret;
1132
1133         INIT_FLIST_HEAD(&list);
1134
1135         memset(&stream, 0, sizeof(stream));
1136         stream.zalloc = Z_NULL;
1137         stream.zfree = Z_NULL;
1138         stream.opaque = Z_NULL;
1139
1140         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1141         if (ret != Z_OK) {
1142                 log_err("fio: failed to init gz stream\n");
1143                 goto err;
1144         }
1145
1146         seq = ++data->log->chunk_seq;
1147
1148         stream.next_in = (void *) data->samples;
1149         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1150
1151         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1152                                 (unsigned long) stream.avail_in, seq,
1153                                 data->log->filename);
1154         do {
1155                 if (c)
1156                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1157                                 (unsigned long) c->len);
1158                 c = get_new_chunk(seq);
1159                 stream.avail_out = GZ_CHUNK;
1160                 stream.next_out = c->buf;
1161                 ret = deflate(&stream, Z_NO_FLUSH);
1162                 if (ret < 0) {
1163                         log_err("fio: deflate log (%d)\n", ret);
1164                         free_chunk(c);
1165                         goto err;
1166                 }
1167
1168                 c->len = GZ_CHUNK - stream.avail_out;
1169                 flist_add_tail(&c->list, &list);
1170                 total += c->len;
1171         } while (stream.avail_in);
1172
1173         stream.next_out = c->buf + c->len;
1174         stream.avail_out = GZ_CHUNK - c->len;
1175
1176         ret = deflate(&stream, Z_FINISH);
1177         if (ret < 0) {
1178                 /*
1179                  * Z_BUF_ERROR is special, it just means we need more
1180                  * output space. We'll handle that below. Treat any other
1181                  * error as fatal.
1182                  */
1183                 if (ret != Z_BUF_ERROR) {
1184                         log_err("fio: deflate log (%d)\n", ret);
1185                         flist_del(&c->list);
1186                         free_chunk(c);
1187                         goto err;
1188                 }
1189         }
1190
1191         total -= c->len;
1192         c->len = GZ_CHUNK - stream.avail_out;
1193         total += c->len;
1194         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1195
1196         if (ret != Z_STREAM_END) {
1197                 do {
1198                         c = get_new_chunk(seq);
1199                         stream.avail_out = GZ_CHUNK;
1200                         stream.next_out = c->buf;
1201                         ret = deflate(&stream, Z_FINISH);
1202                         c->len = GZ_CHUNK - stream.avail_out;
1203                         total += c->len;
1204                         flist_add_tail(&c->list, &list);
1205                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1206                                 (unsigned long) c->len);
1207                 } while (ret != Z_STREAM_END);
1208         }
1209
1210         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1211
1212         ret = deflateEnd(&stream);
1213         if (ret != Z_OK)
1214                 log_err("fio: deflateEnd %d\n", ret);
1215
1216         free(data->samples);
1217
1218         if (!flist_empty(&list)) {
1219                 pthread_mutex_lock(&data->log->chunk_lock);
1220                 flist_splice_tail(&list, &data->log->chunk_list);
1221                 pthread_mutex_unlock(&data->log->chunk_lock);
1222         }
1223
1224         ret = 0;
1225 done:
1226         if (data->free)
1227                 free(data);
1228         return ret;
1229 err:
1230         while (!flist_empty(&list)) {
1231                 c = flist_first_entry(list.next, struct iolog_compress, list);
1232                 flist_del(&c->list);
1233                 free_chunk(c);
1234         }
1235         ret = 1;
1236         goto done;
1237 }
1238
1239 /*
1240  * Invoked from our compress helper thread, when logging would have exceeded
1241  * the specified memory limitation. Compresses the previously stored
1242  * entries.
1243  */
1244 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1245 {
1246         return gz_work(container_of(work, struct iolog_flush_data, work));
1247 }
1248
1249 static int gz_init_worker(struct submit_worker *sw)
1250 {
1251         struct thread_data *td = sw->wq->td;
1252
1253         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1254                 return 0;
1255
1256         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1257                 log_err("gz: failed to set CPU affinity\n");
1258                 return 1;
1259         }
1260
1261         return 0;
1262 }
1263
1264 static struct workqueue_ops log_compress_wq_ops = {
1265         .fn             = gz_work_async,
1266         .init_worker_fn = gz_init_worker,
1267         .nice           = 1,
1268 };
1269
1270 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1271 {
1272         if (!(td->flags & TD_F_COMPRESS_LOG))
1273                 return 0;
1274
1275         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1276         return 0;
1277 }
1278
1279 void iolog_compress_exit(struct thread_data *td)
1280 {
1281         if (!(td->flags & TD_F_COMPRESS_LOG))
1282                 return;
1283
1284         workqueue_exit(&td->log_compress_wq);
1285 }
1286
1287 /*
1288  * Queue work item to compress the existing log entries. We reset the
1289  * current log to a small size, and reference the existing log in the
1290  * data that we queue for compression. Once compression has been done,
1291  * this old log is freed. If called with finish == true, will not return
1292  * until the log compression has completed, and will flush all previous
1293  * logs too
1294  */
1295 static int iolog_flush(struct io_log *log)
1296 {
1297         struct iolog_flush_data *data;
1298
1299         data = malloc(sizeof(*data));
1300         if (!data)
1301                 return 1;
1302
1303         data->log = log;
1304         data->free = false;
1305
1306         while (!flist_empty(&log->io_logs)) {
1307                 struct io_logs *cur_log;
1308
1309                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1310                 flist_del_init(&cur_log->list);
1311
1312                 data->samples = cur_log->log;
1313                 data->nr_samples = cur_log->nr_samples;
1314
1315                 sfree(cur_log);
1316
1317                 gz_work(data);
1318         }
1319
1320         free(data);
1321         return 0;
1322 }
1323
1324 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1325 {
1326         struct iolog_flush_data *data;
1327
1328         data = malloc(sizeof(*data));
1329         if (!data)
1330                 return 1;
1331
1332         data->log = log;
1333
1334         data->samples = cur_log->log;
1335         data->nr_samples = cur_log->nr_samples;
1336         data->free = true;
1337
1338         cur_log->nr_samples = cur_log->max_samples = 0;
1339         cur_log->log = NULL;
1340
1341         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1342         return 0;
1343 }
1344 #else
1345
1346 static int iolog_flush(struct io_log *log)
1347 {
1348         return 1;
1349 }
1350
1351 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1352 {
1353         return 1;
1354 }
1355
1356 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1357 {
1358         return 0;
1359 }
1360
1361 void iolog_compress_exit(struct thread_data *td)
1362 {
1363 }
1364
1365 #endif
1366
1367 struct io_logs *iolog_cur_log(struct io_log *log)
1368 {
1369         if (flist_empty(&log->io_logs))
1370                 return NULL;
1371
1372         return flist_last_entry(&log->io_logs, struct io_logs, list);
1373 }
1374
1375 uint64_t iolog_nr_samples(struct io_log *iolog)
1376 {
1377         struct flist_head *entry;
1378         uint64_t ret = 0;
1379
1380         flist_for_each(entry, &iolog->io_logs) {
1381                 struct io_logs *cur_log;
1382
1383                 cur_log = flist_entry(entry, struct io_logs, list);
1384                 ret += cur_log->nr_samples;
1385         }
1386
1387         return ret;
1388 }
1389
1390 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1391 {
1392         if (log)
1393                 return finish_log(td, log, try);
1394
1395         return 0;
1396 }
1397
1398 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1399 {
1400         int ret;
1401
1402         if (per_unit_log(td->iops_log) != unit_log)
1403                 return 0;
1404
1405         ret = __write_log(td, td->iops_log, try);
1406         if (!ret)
1407                 td->iops_log = NULL;
1408
1409         return ret;
1410 }
1411
1412 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1413 {
1414         int ret;
1415
1416         if (!unit_log)
1417                 return 0;
1418
1419         ret = __write_log(td, td->slat_log, try);
1420         if (!ret)
1421                 td->slat_log = NULL;
1422
1423         return ret;
1424 }
1425
1426 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1427 {
1428         int ret;
1429
1430         if (!unit_log)
1431                 return 0;
1432
1433         ret = __write_log(td, td->clat_log, try);
1434         if (!ret)
1435                 td->clat_log = NULL;
1436
1437         return ret;
1438 }
1439
1440 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1441 {
1442         int ret;
1443
1444         if (!unit_log)
1445                 return 0;
1446
1447         ret = __write_log(td, td->clat_hist_log, try);
1448         if (!ret)
1449                 td->clat_hist_log = NULL;
1450
1451         return ret;
1452 }
1453
1454 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1455 {
1456         int ret;
1457
1458         if (!unit_log)
1459                 return 0;
1460
1461         ret = __write_log(td, td->lat_log, try);
1462         if (!ret)
1463                 td->lat_log = NULL;
1464
1465         return ret;
1466 }
1467
1468 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1469 {
1470         int ret;
1471
1472         if (per_unit_log(td->bw_log) != unit_log)
1473                 return 0;
1474
1475         ret = __write_log(td, td->bw_log, try);
1476         if (!ret)
1477                 td->bw_log = NULL;
1478
1479         return ret;
1480 }
1481
1482 enum {
1483         BW_LOG_MASK     = 1,
1484         LAT_LOG_MASK    = 2,
1485         SLAT_LOG_MASK   = 4,
1486         CLAT_LOG_MASK   = 8,
1487         IOPS_LOG_MASK   = 16,
1488         CLAT_HIST_LOG_MASK = 32,
1489
1490         ALL_LOG_NR      = 6,
1491 };
1492
1493 struct log_type {
1494         unsigned int mask;
1495         int (*fn)(struct thread_data *, int, bool);
1496 };
1497
1498 static struct log_type log_types[] = {
1499         {
1500                 .mask   = BW_LOG_MASK,
1501                 .fn     = write_bandw_log,
1502         },
1503         {
1504                 .mask   = LAT_LOG_MASK,
1505                 .fn     = write_lat_log,
1506         },
1507         {
1508                 .mask   = SLAT_LOG_MASK,
1509                 .fn     = write_slat_log,
1510         },
1511         {
1512                 .mask   = CLAT_LOG_MASK,
1513                 .fn     = write_clat_log,
1514         },
1515         {
1516                 .mask   = IOPS_LOG_MASK,
1517                 .fn     = write_iops_log,
1518         },
1519         {
1520                 .mask   = CLAT_HIST_LOG_MASK,
1521                 .fn     = write_clat_hist_log,
1522         }
1523 };
1524
1525 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1526 {
1527         unsigned int log_mask = 0;
1528         unsigned int log_left = ALL_LOG_NR;
1529         int old_state, i;
1530
1531         old_state = td_bump_runstate(td, TD_FINISHING);
1532
1533         finalize_logs(td, unit_logs);
1534
1535         while (log_left) {
1536                 int prev_log_left = log_left;
1537
1538                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1539                         struct log_type *lt = &log_types[i];
1540                         int ret;
1541
1542                         if (!(log_mask & lt->mask)) {
1543                                 ret = lt->fn(td, log_left != 1, unit_logs);
1544                                 if (!ret) {
1545                                         log_left--;
1546                                         log_mask |= lt->mask;
1547                                 }
1548                         }
1549                 }
1550
1551                 if (prev_log_left == log_left)
1552                         usleep(5000);
1553         }
1554
1555         td_restore_runstate(td, old_state);
1556 }
1557
1558 void fio_writeout_logs(bool unit_logs)
1559 {
1560         struct thread_data *td;
1561         int i;
1562
1563         for_each_td(td, i)
1564                 td_writeout_logs(td, unit_logs);
1565 }