use ARRAY_SIZE() for ignore_error_nr[etype]
[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 #include "blktrace.h"
23
24 static int iolog_flush(struct io_log *log);
25
26 static const char iolog_ver2[] = "fio version 2 iolog";
27
28 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
29 {
30         flist_add_tail(&ipo->list, &td->io_log_list);
31         td->total_io_size += ipo->len;
32 }
33
34 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
35 {
36         if (!td->o.write_iolog_file)
37                 return;
38
39         fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
40                                                 io_ddir_name(io_u->ddir),
41                                                 io_u->offset, io_u->buflen);
42 }
43
44 void log_file(struct thread_data *td, struct fio_file *f,
45               enum file_log_act what)
46 {
47         const char *act[] = { "add", "open", "close" };
48
49         assert(what < 3);
50
51         if (!td->o.write_iolog_file)
52                 return;
53
54
55         /*
56          * this happens on the pre-open/close done before the job starts
57          */
58         if (!td->iolog_f)
59                 return;
60
61         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
62 }
63
64 static void iolog_delay(struct thread_data *td, unsigned long delay)
65 {
66         uint64_t usec = utime_since_now(&td->last_issue);
67         uint64_t this_delay;
68         struct timeval tv;
69
70         if (delay < td->time_offset) {
71                 td->time_offset = 0;
72                 return;
73         }
74
75         delay -= td->time_offset;
76         if (delay < usec)
77                 return;
78
79         delay -= usec;
80
81         fio_gettime(&tv, NULL);
82         while (delay && !td->terminate) {
83                 this_delay = delay;
84                 if (this_delay > 500000)
85                         this_delay = 500000;
86
87                 usec_sleep(td, this_delay);
88                 delay -= this_delay;
89         }
90
91         usec = utime_since_now(&tv);
92         if (usec > delay)
93                 td->time_offset = usec - delay;
94         else
95                 td->time_offset = 0;
96 }
97
98 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
99 {
100         struct fio_file *f;
101         int ret;
102
103         /*
104          * Not a special ipo
105          */
106         if (ipo->ddir != DDIR_INVAL)
107                 return 0;
108
109         f = td->files[ipo->fileno];
110
111         switch (ipo->file_action) {
112         case FIO_LOG_OPEN_FILE:
113                 if (td->o.replay_redirect && fio_file_open(f)) {
114                         dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
115                                         f->file_name);
116                         break;
117                 }
118                 ret = td_io_open_file(td, f);
119                 if (!ret)
120                         break;
121                 td_verror(td, ret, "iolog open file");
122                 return -1;
123         case FIO_LOG_CLOSE_FILE:
124                 td_io_close_file(td, f);
125                 break;
126         case FIO_LOG_UNLINK_FILE:
127                 td_io_unlink_file(td, f);
128                 break;
129         default:
130                 log_err("fio: bad file action %d\n", ipo->file_action);
131                 break;
132         }
133
134         return 1;
135 }
136
137 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
138 {
139         struct io_piece *ipo;
140         unsigned long elapsed;
141
142         while (!flist_empty(&td->io_log_list)) {
143                 int ret;
144
145                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
146                 flist_del(&ipo->list);
147                 remove_trim_entry(td, ipo);
148
149                 ret = ipo_special(td, ipo);
150                 if (ret < 0) {
151                         free(ipo);
152                         break;
153                 } else if (ret > 0) {
154                         free(ipo);
155                         continue;
156                 }
157
158                 io_u->ddir = ipo->ddir;
159                 if (ipo->ddir != DDIR_WAIT) {
160                         io_u->offset = ipo->offset;
161                         io_u->buflen = ipo->len;
162                         io_u->file = td->files[ipo->fileno];
163                         get_file(io_u->file);
164                         dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
165                                                 io_u->buflen, io_u->file->file_name);
166                         if (ipo->delay)
167                                 iolog_delay(td, ipo->delay);
168                 } else {
169                         elapsed = mtime_since_genesis();
170                         if (ipo->delay > elapsed)
171                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
172                 }
173
174                 free(ipo);
175
176                 if (io_u->ddir != DDIR_WAIT)
177                         return 0;
178         }
179
180         td->done = 1;
181         return 1;
182 }
183
184 void prune_io_piece_log(struct thread_data *td)
185 {
186         struct io_piece *ipo;
187         struct rb_node *n;
188
189         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
190                 ipo = rb_entry(n, struct io_piece, rb_node);
191                 rb_erase(n, &td->io_hist_tree);
192                 remove_trim_entry(td, ipo);
193                 td->io_hist_len--;
194                 free(ipo);
195         }
196
197         while (!flist_empty(&td->io_hist_list)) {
198                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
199                 flist_del(&ipo->list);
200                 remove_trim_entry(td, ipo);
201                 td->io_hist_len--;
202                 free(ipo);
203         }
204 }
205
206 /*
207  * log a successful write, so we can unwind the log for verify
208  */
209 void log_io_piece(struct thread_data *td, struct io_u *io_u)
210 {
211         struct rb_node **p, *parent;
212         struct io_piece *ipo, *__ipo;
213
214         ipo = malloc(sizeof(struct io_piece));
215         init_ipo(ipo);
216         ipo->file = io_u->file;
217         ipo->offset = io_u->offset;
218         ipo->len = io_u->buflen;
219         ipo->numberio = io_u->numberio;
220         ipo->flags = IP_F_IN_FLIGHT;
221
222         io_u->ipo = ipo;
223
224         if (io_u_should_trim(td, io_u)) {
225                 flist_add_tail(&ipo->trim_list, &td->trim_list);
226                 td->trim_entries++;
227         }
228
229         /*
230          * We don't need to sort the entries, if:
231          *
232          *      Sequential writes, or
233          *      Random writes that lay out the file as it goes along
234          *
235          * For both these cases, just reading back data in the order we
236          * wrote it out is the fastest.
237          *
238          * One exception is if we don't have a random map AND we are doing
239          * verifies, in that case we need to check for duplicate blocks and
240          * drop the old one, which we rely on the rb insert/lookup for
241          * handling.
242          */
243         if (((!td->o.verifysort) || !td_random(td) || !td->o.overwrite) &&
244               (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
245                 INIT_FLIST_HEAD(&ipo->list);
246                 flist_add_tail(&ipo->list, &td->io_hist_list);
247                 ipo->flags |= IP_F_ONLIST;
248                 td->io_hist_len++;
249                 return;
250         }
251
252         RB_CLEAR_NODE(&ipo->rb_node);
253
254         /*
255          * Sort the entry into the verification list
256          */
257 restart:
258         p = &td->io_hist_tree.rb_node;
259         parent = NULL;
260         while (*p) {
261                 int overlap = 0;
262                 parent = *p;
263
264                 __ipo = rb_entry(parent, struct io_piece, rb_node);
265                 if (ipo->file < __ipo->file)
266                         p = &(*p)->rb_left;
267                 else if (ipo->file > __ipo->file)
268                         p = &(*p)->rb_right;
269                 else if (ipo->offset < __ipo->offset) {
270                         p = &(*p)->rb_left;
271                         overlap = ipo->offset + ipo->len > __ipo->offset;
272                 }
273                 else if (ipo->offset > __ipo->offset) {
274                         p = &(*p)->rb_right;
275                         overlap = __ipo->offset + __ipo->len > ipo->offset;
276                 }
277                 else
278                         overlap = 1;
279
280                 if (overlap) {
281                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
282                                 __ipo->offset, __ipo->len,
283                                 ipo->offset, ipo->len);
284                         td->io_hist_len--;
285                         rb_erase(parent, &td->io_hist_tree);
286                         remove_trim_entry(td, __ipo);
287                         free(__ipo);
288                         goto restart;
289                 }
290         }
291
292         rb_link_node(&ipo->rb_node, parent, p);
293         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
294         ipo->flags |= IP_F_ONRB;
295         td->io_hist_len++;
296 }
297
298 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
299 {
300         struct io_piece *ipo = io_u->ipo;
301
302         if (td->ts.nr_block_infos) {
303                 uint32_t *info = io_u_block_info(td, io_u);
304                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
305                         if (io_u->ddir == DDIR_TRIM)
306                                 *info = BLOCK_INFO_SET_STATE(*info,
307                                                 BLOCK_STATE_TRIM_FAILURE);
308                         else if (io_u->ddir == DDIR_WRITE)
309                                 *info = BLOCK_INFO_SET_STATE(*info,
310                                                 BLOCK_STATE_WRITE_FAILURE);
311                 }
312         }
313
314         if (!ipo)
315                 return;
316
317         if (ipo->flags & IP_F_ONRB)
318                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
319         else if (ipo->flags & IP_F_ONLIST)
320                 flist_del(&ipo->list);
321
322         free(ipo);
323         io_u->ipo = NULL;
324         td->io_hist_len--;
325 }
326
327 void trim_io_piece(struct thread_data *td, const struct io_u *io_u)
328 {
329         struct io_piece *ipo = io_u->ipo;
330
331         if (!ipo)
332                 return;
333
334         ipo->len = io_u->xfer_buflen - io_u->resid;
335 }
336
337 void write_iolog_close(struct thread_data *td)
338 {
339         fflush(td->iolog_f);
340         fclose(td->iolog_f);
341         free(td->iolog_buf);
342         td->iolog_f = NULL;
343         td->iolog_buf = NULL;
344 }
345
346 /*
347  * Read version 2 iolog data. It is enhanced to include per-file logging,
348  * syncs, etc.
349  */
350 static int read_iolog2(struct thread_data *td, FILE *f)
351 {
352         unsigned long long offset;
353         unsigned int bytes;
354         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
355         char *rfname, *fname, *act;
356         char *str, *p;
357         enum fio_ddir rw;
358
359         free_release_files(td);
360
361         /*
362          * Read in the read iolog and store it, reuse the infrastructure
363          * for doing verifications.
364          */
365         str = malloc(4096);
366         rfname = fname = malloc(256+16);
367         act = malloc(256+16);
368
369         reads = writes = waits = 0;
370         while ((p = fgets(str, 4096, f)) != NULL) {
371                 struct io_piece *ipo;
372                 int r;
373
374                 r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
375                                                                         &bytes);
376
377                 if (td->o.replay_redirect)
378                         fname = td->o.replay_redirect;
379
380                 if (r == 4) {
381                         /*
382                          * Check action first
383                          */
384                         if (!strcmp(act, "wait"))
385                                 rw = DDIR_WAIT;
386                         else if (!strcmp(act, "read"))
387                                 rw = DDIR_READ;
388                         else if (!strcmp(act, "write"))
389                                 rw = DDIR_WRITE;
390                         else if (!strcmp(act, "sync"))
391                                 rw = DDIR_SYNC;
392                         else if (!strcmp(act, "datasync"))
393                                 rw = DDIR_DATASYNC;
394                         else if (!strcmp(act, "trim"))
395                                 rw = DDIR_TRIM;
396                         else {
397                                 log_err("fio: bad iolog file action: %s\n",
398                                                                         act);
399                                 continue;
400                         }
401                         fileno = get_fileno(td, fname);
402                 } else if (r == 2) {
403                         rw = DDIR_INVAL;
404                         if (!strcmp(act, "add")) {
405                                 if (td->o.replay_redirect &&
406                                     get_fileno(td, fname) != -1) {
407                                         dprint(FD_FILE, "iolog: ignoring"
408                                                 " re-add of file %s\n", fname);
409                                 } else {
410                                         fileno = add_file(td, fname, 0, 1);
411                                         file_action = FIO_LOG_ADD_FILE;
412                                 }
413                                 continue;
414                         } else if (!strcmp(act, "open")) {
415                                 fileno = get_fileno(td, fname);
416                                 file_action = FIO_LOG_OPEN_FILE;
417                         } else if (!strcmp(act, "close")) {
418                                 fileno = get_fileno(td, fname);
419                                 file_action = FIO_LOG_CLOSE_FILE;
420                         } else {
421                                 log_err("fio: bad iolog file action: %s\n",
422                                                                         act);
423                                 continue;
424                         }
425                 } else {
426                         log_err("bad iolog2: %s\n", p);
427                         continue;
428                 }
429
430                 if (rw == DDIR_READ)
431                         reads++;
432                 else if (rw == DDIR_WRITE) {
433                         /*
434                          * Don't add a write for ro mode
435                          */
436                         if (read_only)
437                                 continue;
438                         writes++;
439                 } else if (rw == DDIR_WAIT) {
440                         if (td->o.no_stall)
441                                 continue;
442                         waits++;
443                 } else if (rw == DDIR_INVAL) {
444                 } else if (!ddir_sync(rw)) {
445                         log_err("bad ddir: %d\n", rw);
446                         continue;
447                 }
448
449                 /*
450                  * Make note of file
451                  */
452                 ipo = malloc(sizeof(*ipo));
453                 init_ipo(ipo);
454                 ipo->ddir = rw;
455                 if (rw == DDIR_WAIT) {
456                         ipo->delay = offset;
457                 } else {
458                         if (td->o.replay_scale)
459                                 ipo->offset = offset / td->o.replay_scale;
460                         else
461                                 ipo->offset = offset;
462                         ipo_bytes_align(td->o.replay_align, ipo);
463
464                         ipo->len = bytes;
465                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw])
466                                 td->o.max_bs[rw] = bytes;
467                         ipo->fileno = fileno;
468                         ipo->file_action = file_action;
469                         td->o.size += bytes;
470                 }
471
472                 queue_io_piece(td, ipo);
473         }
474
475         free(str);
476         free(act);
477         free(rfname);
478
479         if (writes && read_only) {
480                 log_err("fio: <%s> skips replay of %d writes due to"
481                         " read-only\n", td->o.name, writes);
482                 writes = 0;
483         }
484
485         if (!reads && !writes && !waits)
486                 return 1;
487         else if (reads && !writes)
488                 td->o.td_ddir = TD_DDIR_READ;
489         else if (!reads && writes)
490                 td->o.td_ddir = TD_DDIR_WRITE;
491         else
492                 td->o.td_ddir = TD_DDIR_RW;
493
494         return 0;
495 }
496
497 /*
498  * open iolog, check version, and call appropriate parser
499  */
500 static int init_iolog_read(struct thread_data *td)
501 {
502         char buffer[256], *p;
503         FILE *f;
504         int ret;
505
506         f = fopen(td->o.read_iolog_file, "r");
507         if (!f) {
508                 perror("fopen read iolog");
509                 return 1;
510         }
511
512         p = fgets(buffer, sizeof(buffer), f);
513         if (!p) {
514                 td_verror(td, errno, "iolog read");
515                 log_err("fio: unable to read iolog\n");
516                 fclose(f);
517                 return 1;
518         }
519
520         /*
521          * version 2 of the iolog stores a specific string as the
522          * first line, check for that
523          */
524         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
525                 ret = read_iolog2(td, f);
526         else {
527                 log_err("fio: iolog version 1 is no longer supported\n");
528                 ret = 1;
529         }
530
531         fclose(f);
532         return ret;
533 }
534
535 /*
536  * Set up a log for storing io patterns.
537  */
538 static int init_iolog_write(struct thread_data *td)
539 {
540         struct fio_file *ff;
541         FILE *f;
542         unsigned int i;
543
544         f = fopen(td->o.write_iolog_file, "a");
545         if (!f) {
546                 perror("fopen write iolog");
547                 return 1;
548         }
549
550         /*
551          * That's it for writing, setup a log buffer and we're done.
552           */
553         td->iolog_f = f;
554         td->iolog_buf = malloc(8192);
555         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
556
557         /*
558          * write our version line
559          */
560         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
561                 perror("iolog init\n");
562                 return 1;
563         }
564
565         /*
566          * add all known files
567          */
568         for_each_file(td, ff, i)
569                 log_file(td, ff, FIO_LOG_ADD_FILE);
570
571         return 0;
572 }
573
574 int init_iolog(struct thread_data *td)
575 {
576         int ret = 0;
577
578         if (td->o.read_iolog_file) {
579                 int need_swap;
580
581                 /*
582                  * Check if it's a blktrace file and load that if possible.
583                  * Otherwise assume it's a normal log file and load that.
584                  */
585                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
586                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
587                 else
588                         ret = init_iolog_read(td);
589         } else if (td->o.write_iolog_file)
590                 ret = init_iolog_write(td);
591
592         if (ret)
593                 td_verror(td, EINVAL, "failed initializing iolog");
594
595         return ret;
596 }
597
598 void setup_log(struct io_log **log, struct log_params *p,
599                const char *filename)
600 {
601         struct io_log *l;
602         int i;
603         struct io_u_plat_entry *entry;
604         struct flist_head *list;
605
606         l = scalloc(1, sizeof(*l));
607         INIT_FLIST_HEAD(&l->io_logs);
608         l->log_type = p->log_type;
609         l->log_offset = p->log_offset;
610         l->log_gz = p->log_gz;
611         l->log_gz_store = p->log_gz_store;
612         l->avg_msec = p->avg_msec;
613         l->hist_msec = p->hist_msec;
614         l->hist_coarseness = p->hist_coarseness;
615         l->filename = strdup(filename);
616         l->td = p->td;
617
618         /* Initialize histogram lists for each r/w direction,
619          * with initial io_u_plat of all zeros:
620          */
621         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
622                 list = &l->hist_window[i].list;
623                 INIT_FLIST_HEAD(list);
624                 entry = calloc(1, sizeof(struct io_u_plat_entry));
625                 flist_add(&entry->list, list);
626         }
627
628         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
629                 struct io_logs *p;
630
631                 p = calloc(1, sizeof(*l->pending));
632                 p->max_samples = DEF_LOG_ENTRIES;
633                 p->log = calloc(p->max_samples, log_entry_sz(l));
634                 l->pending = p;
635         }
636
637         if (l->log_offset)
638                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
639
640         INIT_FLIST_HEAD(&l->chunk_list);
641
642         if (l->log_gz && !p->td)
643                 l->log_gz = 0;
644         else if (l->log_gz || l->log_gz_store) {
645                 mutex_init_pshared(&l->chunk_lock);
646                 p->td->flags |= TD_F_COMPRESS_LOG;
647         }
648
649         *log = l;
650 }
651
652 #ifdef CONFIG_SETVBUF
653 static void *set_file_buffer(FILE *f)
654 {
655         size_t size = 1048576;
656         void *buf;
657
658         buf = malloc(size);
659         setvbuf(f, buf, _IOFBF, size);
660         return buf;
661 }
662
663 static void clear_file_buffer(void *buf)
664 {
665         free(buf);
666 }
667 #else
668 static void *set_file_buffer(FILE *f)
669 {
670         return NULL;
671 }
672
673 static void clear_file_buffer(void *buf)
674 {
675 }
676 #endif
677
678 void free_log(struct io_log *log)
679 {
680         while (!flist_empty(&log->io_logs)) {
681                 struct io_logs *cur_log;
682
683                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
684                 flist_del_init(&cur_log->list);
685                 free(cur_log->log);
686                 sfree(cur_log);
687         }
688
689         if (log->pending) {
690                 free(log->pending->log);
691                 free(log->pending);
692                 log->pending = NULL;
693         }
694
695         free(log->pending);
696         free(log->filename);
697         sfree(log);
698 }
699
700 unsigned long hist_sum(int j, int stride, unsigned int *io_u_plat,
701                 unsigned int *io_u_plat_last)
702 {
703         unsigned long sum;
704         int k;
705
706         if (io_u_plat_last) {
707                 for (k = sum = 0; k < stride; k++)
708                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
709         } else {
710                 for (k = sum = 0; k < stride; k++)
711                         sum += io_u_plat[j + k];
712         }
713
714         return sum;
715 }
716
717 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
718                                uint64_t sample_size)
719 {
720         struct io_sample *s;
721         int log_offset;
722         uint64_t i, j, nr_samples;
723         struct io_u_plat_entry *entry, *entry_before;
724         unsigned int *io_u_plat;
725         unsigned int *io_u_plat_before;
726
727         int stride = 1 << hist_coarseness;
728         
729         if (!sample_size)
730                 return;
731
732         s = __get_sample(samples, 0, 0);
733         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
734
735         nr_samples = sample_size / __log_entry_sz(log_offset);
736
737         for (i = 0; i < nr_samples; i++) {
738                 s = __get_sample(samples, log_offset, i);
739
740                 entry = s->data.plat_entry;
741                 io_u_plat = entry->io_u_plat;
742
743                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
744                 io_u_plat_before = entry_before->io_u_plat;
745
746                 fprintf(f, "%lu, %u, %u, ", (unsigned long) s->time,
747                                                 io_sample_ddir(s), s->bs);
748                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
749                         fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat,
750                                                 io_u_plat_before));
751                 }
752                 fprintf(f, "%lu\n", (unsigned long)
753                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
754                                         io_u_plat_before));
755
756                 flist_del(&entry_before->list);
757                 free(entry_before);
758         }
759 }
760
761 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
762 {
763         struct io_sample *s;
764         int log_offset;
765         uint64_t i, nr_samples;
766
767         if (!sample_size)
768                 return;
769
770         s = __get_sample(samples, 0, 0);
771         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
772
773         nr_samples = sample_size / __log_entry_sz(log_offset);
774
775         for (i = 0; i < nr_samples; i++) {
776                 s = __get_sample(samples, log_offset, i);
777
778                 if (!log_offset) {
779                         fprintf(f, "%lu, %" PRId64 ", %u, %u\n",
780                                         (unsigned long) s->time,
781                                         s->data.val,
782                                         io_sample_ddir(s), s->bs);
783                 } else {
784                         struct io_sample_offset *so = (void *) s;
785
786                         fprintf(f, "%lu, %" PRId64 ", %u, %u, %llu\n",
787                                         (unsigned long) s->time,
788                                         s->data.val,
789                                         io_sample_ddir(s), s->bs,
790                                         (unsigned long long) so->offset);
791                 }
792         }
793 }
794
795 #ifdef CONFIG_ZLIB
796
797 struct iolog_flush_data {
798         struct workqueue_work work;
799         struct io_log *log;
800         void *samples;
801         uint32_t nr_samples;
802         bool free;
803 };
804
805 #define GZ_CHUNK        131072
806
807 static struct iolog_compress *get_new_chunk(unsigned int seq)
808 {
809         struct iolog_compress *c;
810
811         c = malloc(sizeof(*c));
812         INIT_FLIST_HEAD(&c->list);
813         c->buf = malloc(GZ_CHUNK);
814         c->len = 0;
815         c->seq = seq;
816         return c;
817 }
818
819 static void free_chunk(struct iolog_compress *ic)
820 {
821         free(ic->buf);
822         free(ic);
823 }
824
825 static int z_stream_init(z_stream *stream, int gz_hdr)
826 {
827         int wbits = 15;
828
829         memset(stream, 0, sizeof(*stream));
830         stream->zalloc = Z_NULL;
831         stream->zfree = Z_NULL;
832         stream->opaque = Z_NULL;
833         stream->next_in = Z_NULL;
834
835         /*
836          * zlib magic - add 32 for auto-detection of gz header or not,
837          * if we decide to store files in a gzip friendly format.
838          */
839         if (gz_hdr)
840                 wbits += 32;
841
842         if (inflateInit2(stream, wbits) != Z_OK)
843                 return 1;
844
845         return 0;
846 }
847
848 struct inflate_chunk_iter {
849         unsigned int seq;
850         int err;
851         void *buf;
852         size_t buf_size;
853         size_t buf_used;
854         size_t chunk_sz;
855 };
856
857 static void finish_chunk(z_stream *stream, FILE *f,
858                          struct inflate_chunk_iter *iter)
859 {
860         int ret;
861
862         ret = inflateEnd(stream);
863         if (ret != Z_OK)
864                 log_err("fio: failed to end log inflation seq %d (%d)\n",
865                                 iter->seq, ret);
866
867         flush_samples(f, iter->buf, iter->buf_used);
868         free(iter->buf);
869         iter->buf = NULL;
870         iter->buf_size = iter->buf_used = 0;
871 }
872
873 /*
874  * Iterative chunk inflation. Handles cases where we cross into a new
875  * sequence, doing flush finish of previous chunk if needed.
876  */
877 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
878                             z_stream *stream, struct inflate_chunk_iter *iter)
879 {
880         size_t ret;
881
882         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
883                                 (unsigned long) ic->len, ic->seq);
884
885         if (ic->seq != iter->seq) {
886                 if (iter->seq)
887                         finish_chunk(stream, f, iter);
888
889                 z_stream_init(stream, gz_hdr);
890                 iter->seq = ic->seq;
891         }
892
893         stream->avail_in = ic->len;
894         stream->next_in = ic->buf;
895
896         if (!iter->buf_size) {
897                 iter->buf_size = iter->chunk_sz;
898                 iter->buf = malloc(iter->buf_size);
899         }
900
901         while (stream->avail_in) {
902                 size_t this_out = iter->buf_size - iter->buf_used;
903                 int err;
904
905                 stream->avail_out = this_out;
906                 stream->next_out = iter->buf + iter->buf_used;
907
908                 err = inflate(stream, Z_NO_FLUSH);
909                 if (err < 0) {
910                         log_err("fio: failed inflating log: %d\n", err);
911                         iter->err = err;
912                         break;
913                 }
914
915                 iter->buf_used += this_out - stream->avail_out;
916
917                 if (!stream->avail_out) {
918                         iter->buf_size += iter->chunk_sz;
919                         iter->buf = realloc(iter->buf, iter->buf_size);
920                         continue;
921                 }
922
923                 if (err == Z_STREAM_END)
924                         break;
925         }
926
927         ret = (void *) stream->next_in - ic->buf;
928
929         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
930
931         return ret;
932 }
933
934 /*
935  * Inflate stored compressed chunks, or write them directly to the log
936  * file if so instructed.
937  */
938 static int inflate_gz_chunks(struct io_log *log, FILE *f)
939 {
940         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
941         z_stream stream;
942
943         while (!flist_empty(&log->chunk_list)) {
944                 struct iolog_compress *ic;
945
946                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
947                 flist_del(&ic->list);
948
949                 if (log->log_gz_store) {
950                         size_t ret;
951
952                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
953                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
954
955                         ret = fwrite(ic->buf, ic->len, 1, f);
956                         if (ret != 1 || ferror(f)) {
957                                 iter.err = errno;
958                                 log_err("fio: error writing compressed log\n");
959                         }
960                 } else
961                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
962
963                 free_chunk(ic);
964         }
965
966         if (iter.seq) {
967                 finish_chunk(&stream, f, &iter);
968                 free(iter.buf);
969         }
970
971         return iter.err;
972 }
973
974 /*
975  * Open compressed log file and decompress the stored chunks and
976  * write them to stdout. The chunks are stored sequentially in the
977  * file, so we iterate over them and do them one-by-one.
978  */
979 int iolog_file_inflate(const char *file)
980 {
981         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
982         struct iolog_compress ic;
983         z_stream stream;
984         struct stat sb;
985         ssize_t ret;
986         size_t total;
987         void *buf;
988         FILE *f;
989
990         f = fopen(file, "r");
991         if (!f) {
992                 perror("fopen");
993                 return 1;
994         }
995
996         if (stat(file, &sb) < 0) {
997                 fclose(f);
998                 perror("stat");
999                 return 1;
1000         }
1001
1002         ic.buf = buf = malloc(sb.st_size);
1003         ic.len = sb.st_size;
1004         ic.seq = 1;
1005
1006         ret = fread(ic.buf, ic.len, 1, f);
1007         if (ret < 0) {
1008                 perror("fread");
1009                 fclose(f);
1010                 free(buf);
1011                 return 1;
1012         } else if (ret != 1) {
1013                 log_err("fio: short read on reading log\n");
1014                 fclose(f);
1015                 free(buf);
1016                 return 1;
1017         }
1018
1019         fclose(f);
1020
1021         /*
1022          * Each chunk will return Z_STREAM_END. We don't know how many
1023          * chunks are in the file, so we just keep looping and incrementing
1024          * the sequence number until we have consumed the whole compressed
1025          * file.
1026          */
1027         total = ic.len;
1028         do {
1029                 size_t iret;
1030
1031                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1032                 total -= iret;
1033                 if (!total)
1034                         break;
1035                 if (iter.err)
1036                         break;
1037
1038                 ic.seq++;
1039                 ic.len -= iret;
1040                 ic.buf += iret;
1041         } while (1);
1042
1043         if (iter.seq) {
1044                 finish_chunk(&stream, stdout, &iter);
1045                 free(iter.buf);
1046         }
1047
1048         free(buf);
1049         return iter.err;
1050 }
1051
1052 #else
1053
1054 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1055 {
1056         return 0;
1057 }
1058
1059 int iolog_file_inflate(const char *file)
1060 {
1061         log_err("fio: log inflation not possible without zlib\n");
1062         return 1;
1063 }
1064
1065 #endif
1066
1067 void flush_log(struct io_log *log, bool do_append)
1068 {
1069         void *buf;
1070         FILE *f;
1071
1072         if (!do_append)
1073                 f = fopen(log->filename, "w");
1074         else
1075                 f = fopen(log->filename, "a");
1076         if (!f) {
1077                 perror("fopen log");
1078                 return;
1079         }
1080
1081         buf = set_file_buffer(f);
1082
1083         inflate_gz_chunks(log, f);
1084
1085         while (!flist_empty(&log->io_logs)) {
1086                 struct io_logs *cur_log;
1087
1088                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1089                 flist_del_init(&cur_log->list);
1090                 
1091                 if (log->td && log == log->td->clat_hist_log)
1092                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1093                                            log_sample_sz(log, cur_log));
1094                 else
1095                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1096                 
1097                 sfree(cur_log);
1098         }
1099
1100         fclose(f);
1101         clear_file_buffer(buf);
1102 }
1103
1104 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1105 {
1106         if (td->flags & TD_F_COMPRESS_LOG)
1107                 iolog_flush(log);
1108
1109         if (trylock) {
1110                 if (fio_trylock_file(log->filename))
1111                         return 1;
1112         } else
1113                 fio_lock_file(log->filename);
1114
1115         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1116                 fio_send_iolog(td, log, log->filename);
1117         else
1118                 flush_log(log, !td->o.per_job_logs);
1119
1120         fio_unlock_file(log->filename);
1121         free_log(log);
1122         return 0;
1123 }
1124
1125 size_t log_chunk_sizes(struct io_log *log)
1126 {
1127         struct flist_head *entry;
1128         size_t ret;
1129
1130         if (flist_empty(&log->chunk_list))
1131                 return 0;
1132
1133         ret = 0;
1134         pthread_mutex_lock(&log->chunk_lock);
1135         flist_for_each(entry, &log->chunk_list) {
1136                 struct iolog_compress *c;
1137
1138                 c = flist_entry(entry, struct iolog_compress, list);
1139                 ret += c->len;
1140         }
1141         pthread_mutex_unlock(&log->chunk_lock);
1142         return ret;
1143 }
1144
1145 #ifdef CONFIG_ZLIB
1146
1147 static int gz_work(struct iolog_flush_data *data)
1148 {
1149         struct iolog_compress *c = NULL;
1150         struct flist_head list;
1151         unsigned int seq;
1152         z_stream stream;
1153         size_t total = 0;
1154         int ret;
1155
1156         INIT_FLIST_HEAD(&list);
1157
1158         memset(&stream, 0, sizeof(stream));
1159         stream.zalloc = Z_NULL;
1160         stream.zfree = Z_NULL;
1161         stream.opaque = Z_NULL;
1162
1163         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1164         if (ret != Z_OK) {
1165                 log_err("fio: failed to init gz stream\n");
1166                 goto err;
1167         }
1168
1169         seq = ++data->log->chunk_seq;
1170
1171         stream.next_in = (void *) data->samples;
1172         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1173
1174         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1175                                 (unsigned long) stream.avail_in, seq,
1176                                 data->log->filename);
1177         do {
1178                 if (c)
1179                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1180                                 (unsigned long) c->len);
1181                 c = get_new_chunk(seq);
1182                 stream.avail_out = GZ_CHUNK;
1183                 stream.next_out = c->buf;
1184                 ret = deflate(&stream, Z_NO_FLUSH);
1185                 if (ret < 0) {
1186                         log_err("fio: deflate log (%d)\n", ret);
1187                         free_chunk(c);
1188                         goto err;
1189                 }
1190
1191                 c->len = GZ_CHUNK - stream.avail_out;
1192                 flist_add_tail(&c->list, &list);
1193                 total += c->len;
1194         } while (stream.avail_in);
1195
1196         stream.next_out = c->buf + c->len;
1197         stream.avail_out = GZ_CHUNK - c->len;
1198
1199         ret = deflate(&stream, Z_FINISH);
1200         if (ret < 0) {
1201                 /*
1202                  * Z_BUF_ERROR is special, it just means we need more
1203                  * output space. We'll handle that below. Treat any other
1204                  * error as fatal.
1205                  */
1206                 if (ret != Z_BUF_ERROR) {
1207                         log_err("fio: deflate log (%d)\n", ret);
1208                         flist_del(&c->list);
1209                         free_chunk(c);
1210                         goto err;
1211                 }
1212         }
1213
1214         total -= c->len;
1215         c->len = GZ_CHUNK - stream.avail_out;
1216         total += c->len;
1217         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1218
1219         if (ret != Z_STREAM_END) {
1220                 do {
1221                         c = get_new_chunk(seq);
1222                         stream.avail_out = GZ_CHUNK;
1223                         stream.next_out = c->buf;
1224                         ret = deflate(&stream, Z_FINISH);
1225                         c->len = GZ_CHUNK - stream.avail_out;
1226                         total += c->len;
1227                         flist_add_tail(&c->list, &list);
1228                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1229                                 (unsigned long) c->len);
1230                 } while (ret != Z_STREAM_END);
1231         }
1232
1233         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1234
1235         ret = deflateEnd(&stream);
1236         if (ret != Z_OK)
1237                 log_err("fio: deflateEnd %d\n", ret);
1238
1239         free(data->samples);
1240
1241         if (!flist_empty(&list)) {
1242                 pthread_mutex_lock(&data->log->chunk_lock);
1243                 flist_splice_tail(&list, &data->log->chunk_list);
1244                 pthread_mutex_unlock(&data->log->chunk_lock);
1245         }
1246
1247         ret = 0;
1248 done:
1249         if (data->free)
1250                 free(data);
1251         return ret;
1252 err:
1253         while (!flist_empty(&list)) {
1254                 c = flist_first_entry(list.next, struct iolog_compress, list);
1255                 flist_del(&c->list);
1256                 free_chunk(c);
1257         }
1258         ret = 1;
1259         goto done;
1260 }
1261
1262 /*
1263  * Invoked from our compress helper thread, when logging would have exceeded
1264  * the specified memory limitation. Compresses the previously stored
1265  * entries.
1266  */
1267 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1268 {
1269         return gz_work(container_of(work, struct iolog_flush_data, work));
1270 }
1271
1272 static int gz_init_worker(struct submit_worker *sw)
1273 {
1274         struct thread_data *td = sw->wq->td;
1275
1276         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1277                 return 0;
1278
1279         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1280                 log_err("gz: failed to set CPU affinity\n");
1281                 return 1;
1282         }
1283
1284         return 0;
1285 }
1286
1287 static struct workqueue_ops log_compress_wq_ops = {
1288         .fn             = gz_work_async,
1289         .init_worker_fn = gz_init_worker,
1290         .nice           = 1,
1291 };
1292
1293 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1294 {
1295         if (!(td->flags & TD_F_COMPRESS_LOG))
1296                 return 0;
1297
1298         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1299         return 0;
1300 }
1301
1302 void iolog_compress_exit(struct thread_data *td)
1303 {
1304         if (!(td->flags & TD_F_COMPRESS_LOG))
1305                 return;
1306
1307         workqueue_exit(&td->log_compress_wq);
1308 }
1309
1310 /*
1311  * Queue work item to compress the existing log entries. We reset the
1312  * current log to a small size, and reference the existing log in the
1313  * data that we queue for compression. Once compression has been done,
1314  * this old log is freed. If called with finish == true, will not return
1315  * until the log compression has completed, and will flush all previous
1316  * logs too
1317  */
1318 static int iolog_flush(struct io_log *log)
1319 {
1320         struct iolog_flush_data *data;
1321
1322         data = malloc(sizeof(*data));
1323         if (!data)
1324                 return 1;
1325
1326         data->log = log;
1327         data->free = false;
1328
1329         while (!flist_empty(&log->io_logs)) {
1330                 struct io_logs *cur_log;
1331
1332                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1333                 flist_del_init(&cur_log->list);
1334
1335                 data->samples = cur_log->log;
1336                 data->nr_samples = cur_log->nr_samples;
1337
1338                 sfree(cur_log);
1339
1340                 gz_work(data);
1341         }
1342
1343         free(data);
1344         return 0;
1345 }
1346
1347 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1348 {
1349         struct iolog_flush_data *data;
1350
1351         data = malloc(sizeof(*data));
1352         if (!data)
1353                 return 1;
1354
1355         data->log = log;
1356
1357         data->samples = cur_log->log;
1358         data->nr_samples = cur_log->nr_samples;
1359         data->free = true;
1360
1361         cur_log->nr_samples = cur_log->max_samples = 0;
1362         cur_log->log = NULL;
1363
1364         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1365         return 0;
1366 }
1367 #else
1368
1369 static int iolog_flush(struct io_log *log)
1370 {
1371         return 1;
1372 }
1373
1374 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1375 {
1376         return 1;
1377 }
1378
1379 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1380 {
1381         return 0;
1382 }
1383
1384 void iolog_compress_exit(struct thread_data *td)
1385 {
1386 }
1387
1388 #endif
1389
1390 struct io_logs *iolog_cur_log(struct io_log *log)
1391 {
1392         if (flist_empty(&log->io_logs))
1393                 return NULL;
1394
1395         return flist_last_entry(&log->io_logs, struct io_logs, list);
1396 }
1397
1398 uint64_t iolog_nr_samples(struct io_log *iolog)
1399 {
1400         struct flist_head *entry;
1401         uint64_t ret = 0;
1402
1403         flist_for_each(entry, &iolog->io_logs) {
1404                 struct io_logs *cur_log;
1405
1406                 cur_log = flist_entry(entry, struct io_logs, list);
1407                 ret += cur_log->nr_samples;
1408         }
1409
1410         return ret;
1411 }
1412
1413 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1414 {
1415         if (log)
1416                 return finish_log(td, log, try);
1417
1418         return 0;
1419 }
1420
1421 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1422 {
1423         int ret;
1424
1425         if (per_unit_log(td->iops_log) != unit_log)
1426                 return 0;
1427
1428         ret = __write_log(td, td->iops_log, try);
1429         if (!ret)
1430                 td->iops_log = NULL;
1431
1432         return ret;
1433 }
1434
1435 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1436 {
1437         int ret;
1438
1439         if (!unit_log)
1440                 return 0;
1441
1442         ret = __write_log(td, td->slat_log, try);
1443         if (!ret)
1444                 td->slat_log = NULL;
1445
1446         return ret;
1447 }
1448
1449 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1450 {
1451         int ret;
1452
1453         if (!unit_log)
1454                 return 0;
1455
1456         ret = __write_log(td, td->clat_log, try);
1457         if (!ret)
1458                 td->clat_log = NULL;
1459
1460         return ret;
1461 }
1462
1463 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1464 {
1465         int ret;
1466
1467         if (!unit_log)
1468                 return 0;
1469
1470         ret = __write_log(td, td->clat_hist_log, try);
1471         if (!ret)
1472                 td->clat_hist_log = NULL;
1473
1474         return ret;
1475 }
1476
1477 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1478 {
1479         int ret;
1480
1481         if (!unit_log)
1482                 return 0;
1483
1484         ret = __write_log(td, td->lat_log, try);
1485         if (!ret)
1486                 td->lat_log = NULL;
1487
1488         return ret;
1489 }
1490
1491 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1492 {
1493         int ret;
1494
1495         if (per_unit_log(td->bw_log) != unit_log)
1496                 return 0;
1497
1498         ret = __write_log(td, td->bw_log, try);
1499         if (!ret)
1500                 td->bw_log = NULL;
1501
1502         return ret;
1503 }
1504
1505 enum {
1506         BW_LOG_MASK     = 1,
1507         LAT_LOG_MASK    = 2,
1508         SLAT_LOG_MASK   = 4,
1509         CLAT_LOG_MASK   = 8,
1510         IOPS_LOG_MASK   = 16,
1511         CLAT_HIST_LOG_MASK = 32,
1512
1513         ALL_LOG_NR      = 6,
1514 };
1515
1516 struct log_type {
1517         unsigned int mask;
1518         int (*fn)(struct thread_data *, int, bool);
1519 };
1520
1521 static struct log_type log_types[] = {
1522         {
1523                 .mask   = BW_LOG_MASK,
1524                 .fn     = write_bandw_log,
1525         },
1526         {
1527                 .mask   = LAT_LOG_MASK,
1528                 .fn     = write_lat_log,
1529         },
1530         {
1531                 .mask   = SLAT_LOG_MASK,
1532                 .fn     = write_slat_log,
1533         },
1534         {
1535                 .mask   = CLAT_LOG_MASK,
1536                 .fn     = write_clat_log,
1537         },
1538         {
1539                 .mask   = IOPS_LOG_MASK,
1540                 .fn     = write_iops_log,
1541         },
1542         {
1543                 .mask   = CLAT_HIST_LOG_MASK,
1544                 .fn     = write_clat_hist_log,
1545         }
1546 };
1547
1548 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1549 {
1550         unsigned int log_mask = 0;
1551         unsigned int log_left = ALL_LOG_NR;
1552         int old_state, i;
1553
1554         old_state = td_bump_runstate(td, TD_FINISHING);
1555
1556         finalize_logs(td, unit_logs);
1557
1558         while (log_left) {
1559                 int prev_log_left = log_left;
1560
1561                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1562                         struct log_type *lt = &log_types[i];
1563                         int ret;
1564
1565                         if (!(log_mask & lt->mask)) {
1566                                 ret = lt->fn(td, log_left != 1, unit_logs);
1567                                 if (!ret) {
1568                                         log_left--;
1569                                         log_mask |= lt->mask;
1570                                 }
1571                         }
1572                 }
1573
1574                 if (prev_log_left == log_left)
1575                         usleep(5000);
1576         }
1577
1578         td_restore_runstate(td, old_state);
1579 }
1580
1581 void fio_writeout_logs(bool unit_logs)
1582 {
1583         struct thread_data *td;
1584         int i;
1585
1586         for_each_td(td, i)
1587                 td_writeout_logs(td, unit_logs);
1588 }