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