iolog: enable replay_redirect on iolog replay
[fio.git] / iolog.c
1 /*
2  * Code related to writing an iolog of what a thread is doing, and to
3  * later read that back and replay
4  */
5 #include <stdio.h>
6 #include <stdlib.h>
7 #include <libgen.h>
8 #include <assert.h>
9 #include <sys/types.h>
10 #include <sys/stat.h>
11 #include <unistd.h>
12 #ifdef CONFIG_ZLIB
13 #include <zlib.h>
14 #endif
15
16 #include "flist.h"
17 #include "fio.h"
18 #include "verify.h"
19 #include "trim.h"
20 #include "filelock.h"
21 #include "smalloc.h"
22
23 static int iolog_flush(struct io_log *log);
24
25 static const char iolog_ver2[] = "fio version 2 iolog";
26
27 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
28 {
29         flist_add_tail(&ipo->list, &td->io_log_list);
30         td->total_io_size += ipo->len;
31 }
32
33 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
34 {
35         if (!td->o.write_iolog_file)
36                 return;
37
38         fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
39                                                 io_ddir_name(io_u->ddir),
40                                                 io_u->offset, io_u->buflen);
41 }
42
43 void log_file(struct thread_data *td, struct fio_file *f,
44               enum file_log_act what)
45 {
46         const char *act[] = { "add", "open", "close" };
47
48         assert(what < 3);
49
50         if (!td->o.write_iolog_file)
51                 return;
52
53
54         /*
55          * this happens on the pre-open/close done before the job starts
56          */
57         if (!td->iolog_f)
58                 return;
59
60         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
61 }
62
63 static void iolog_delay(struct thread_data *td, unsigned long delay)
64 {
65         uint64_t usec = utime_since_now(&td->last_issue);
66         uint64_t this_delay;
67         struct timeval tv;
68
69         if (delay < td->time_offset) {
70                 td->time_offset = 0;
71                 return;
72         }
73
74         delay -= td->time_offset;
75         if (delay < usec)
76                 return;
77
78         delay -= usec;
79
80         fio_gettime(&tv, NULL);
81         while (delay && !td->terminate) {
82                 this_delay = delay;
83                 if (this_delay > 500000)
84                         this_delay = 500000;
85
86                 usec_sleep(td, this_delay);
87                 delay -= this_delay;
88         }
89
90         usec = utime_since_now(&tv);
91         if (usec > delay)
92                 td->time_offset = usec - delay;
93         else
94                 td->time_offset = 0;
95 }
96
97 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
98 {
99         struct fio_file *f;
100         int ret;
101
102         /*
103          * Not a special ipo
104          */
105         if (ipo->ddir != DDIR_INVAL)
106                 return 0;
107
108         f = td->files[ipo->fileno];
109
110         switch (ipo->file_action) {
111         case FIO_LOG_OPEN_FILE:
112                 ret = td_io_open_file(td, f);
113                 if (!ret)
114                         break;
115                 td_verror(td, ret, "iolog open file");
116                 return -1;
117         case FIO_LOG_CLOSE_FILE:
118                 td_io_close_file(td, f);
119                 break;
120         case FIO_LOG_UNLINK_FILE:
121                 td_io_unlink_file(td, f);
122                 break;
123         default:
124                 log_err("fio: bad file action %d\n", ipo->file_action);
125                 break;
126         }
127
128         return 1;
129 }
130
131 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
132 {
133         struct io_piece *ipo;
134         unsigned long elapsed;
135
136         while (!flist_empty(&td->io_log_list)) {
137                 int ret;
138
139                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
140                 flist_del(&ipo->list);
141                 remove_trim_entry(td, ipo);
142
143                 ret = ipo_special(td, ipo);
144                 if (ret < 0) {
145                         free(ipo);
146                         break;
147                 } else if (ret > 0) {
148                         free(ipo);
149                         continue;
150                 }
151
152                 io_u->ddir = ipo->ddir;
153                 if (ipo->ddir != DDIR_WAIT) {
154                         io_u->offset = ipo->offset;
155                         io_u->buflen = ipo->len;
156                         io_u->file = td->files[ipo->fileno];
157                         get_file(io_u->file);
158                         dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
159                                                 io_u->buflen, io_u->file->file_name);
160                         if (ipo->delay)
161                                 iolog_delay(td, ipo->delay);
162                 } else {
163                         elapsed = mtime_since_genesis();
164                         if (ipo->delay > elapsed)
165                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
166                 }
167
168                 free(ipo);
169
170                 if (io_u->ddir != DDIR_WAIT)
171                         return 0;
172         }
173
174         td->done = 1;
175         return 1;
176 }
177
178 void prune_io_piece_log(struct thread_data *td)
179 {
180         struct io_piece *ipo;
181         struct rb_node *n;
182
183         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
184                 ipo = rb_entry(n, struct io_piece, rb_node);
185                 rb_erase(n, &td->io_hist_tree);
186                 remove_trim_entry(td, ipo);
187                 td->io_hist_len--;
188                 free(ipo);
189         }
190
191         while (!flist_empty(&td->io_hist_list)) {
192                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
193                 flist_del(&ipo->list);
194                 remove_trim_entry(td, ipo);
195                 td->io_hist_len--;
196                 free(ipo);
197         }
198 }
199
200 /*
201  * log a successful write, so we can unwind the log for verify
202  */
203 void log_io_piece(struct thread_data *td, struct io_u *io_u)
204 {
205         struct rb_node **p, *parent;
206         struct io_piece *ipo, *__ipo;
207
208         ipo = malloc(sizeof(struct io_piece));
209         init_ipo(ipo);
210         ipo->file = io_u->file;
211         ipo->offset = io_u->offset;
212         ipo->len = io_u->buflen;
213         ipo->numberio = io_u->numberio;
214         ipo->flags = IP_F_IN_FLIGHT;
215
216         io_u->ipo = ipo;
217
218         if (io_u_should_trim(td, io_u)) {
219                 flist_add_tail(&ipo->trim_list, &td->trim_list);
220                 td->trim_entries++;
221         }
222
223         /*
224          * We don't need to sort the entries, if:
225          *
226          *      Sequential writes, or
227          *      Random writes that lay out the file as it goes along
228          *
229          * For both these cases, just reading back data in the order we
230          * wrote it out is the fastest.
231          *
232          * One exception is if we don't have a random map AND we are doing
233          * verifies, in that case we need to check for duplicate blocks and
234          * drop the old one, which we rely on the rb insert/lookup for
235          * handling.
236          */
237         if (((!td->o.verifysort) || !td_random(td) || !td->o.overwrite) &&
238               (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
239                 INIT_FLIST_HEAD(&ipo->list);
240                 flist_add_tail(&ipo->list, &td->io_hist_list);
241                 ipo->flags |= IP_F_ONLIST;
242                 td->io_hist_len++;
243                 return;
244         }
245
246         RB_CLEAR_NODE(&ipo->rb_node);
247
248         /*
249          * Sort the entry into the verification list
250          */
251 restart:
252         p = &td->io_hist_tree.rb_node;
253         parent = NULL;
254         while (*p) {
255                 int overlap = 0;
256                 parent = *p;
257
258                 __ipo = rb_entry(parent, struct io_piece, rb_node);
259                 if (ipo->file < __ipo->file)
260                         p = &(*p)->rb_left;
261                 else if (ipo->file > __ipo->file)
262                         p = &(*p)->rb_right;
263                 else if (ipo->offset < __ipo->offset) {
264                         p = &(*p)->rb_left;
265                         overlap = ipo->offset + ipo->len > __ipo->offset;
266                 }
267                 else if (ipo->offset > __ipo->offset) {
268                         p = &(*p)->rb_right;
269                         overlap = __ipo->offset + __ipo->len > ipo->offset;
270                 }
271                 else
272                         overlap = 1;
273
274                 if (overlap) {
275                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu",
276                                 __ipo->offset, __ipo->len,
277                                 ipo->offset, ipo->len);
278                         td->io_hist_len--;
279                         rb_erase(parent, &td->io_hist_tree);
280                         remove_trim_entry(td, __ipo);
281                         free(__ipo);
282                         goto restart;
283                 }
284         }
285
286         rb_link_node(&ipo->rb_node, parent, p);
287         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
288         ipo->flags |= IP_F_ONRB;
289         td->io_hist_len++;
290 }
291
292 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
293 {
294         struct io_piece *ipo = io_u->ipo;
295
296         if (td->ts.nr_block_infos) {
297                 uint32_t *info = io_u_block_info(td, io_u);
298                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
299                         if (io_u->ddir == DDIR_TRIM)
300                                 *info = BLOCK_INFO_SET_STATE(*info,
301                                                 BLOCK_STATE_TRIM_FAILURE);
302                         else if (io_u->ddir == DDIR_WRITE)
303                                 *info = BLOCK_INFO_SET_STATE(*info,
304                                                 BLOCK_STATE_WRITE_FAILURE);
305                 }
306         }
307
308         if (!ipo)
309                 return;
310
311         if (ipo->flags & IP_F_ONRB)
312                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
313         else if (ipo->flags & IP_F_ONLIST)
314                 flist_del(&ipo->list);
315
316         free(ipo);
317         io_u->ipo = NULL;
318         td->io_hist_len--;
319 }
320
321 void trim_io_piece(struct thread_data *td, const struct io_u *io_u)
322 {
323         struct io_piece *ipo = io_u->ipo;
324
325         if (!ipo)
326                 return;
327
328         ipo->len = io_u->xfer_buflen - io_u->resid;
329 }
330
331 void write_iolog_close(struct thread_data *td)
332 {
333         fflush(td->iolog_f);
334         fclose(td->iolog_f);
335         free(td->iolog_buf);
336         td->iolog_f = NULL;
337         td->iolog_buf = NULL;
338 }
339
340 /*
341  * Read version 2 iolog data. It is enhanced to include per-file logging,
342  * syncs, etc.
343  */
344 static int read_iolog2(struct thread_data *td, FILE *f)
345 {
346         unsigned long long offset;
347         unsigned int bytes;
348         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
349         char *rfname, *fname, *act;
350         char *str, *p;
351         enum fio_ddir rw;
352
353         free_release_files(td);
354
355         /*
356          * Read in the read iolog and store it, reuse the infrastructure
357          * for doing verifications.
358          */
359         str = malloc(4096);
360         rfname = fname = malloc(256+16);
361         act = malloc(256+16);
362
363         reads = writes = waits = 0;
364         while ((p = fgets(str, 4096, f)) != NULL) {
365                 struct io_piece *ipo;
366                 int r;
367
368                 r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
369                                                                         &bytes);
370
371                 if (td->o.replay_redirect)
372                         fname = td->o.replay_redirect;
373
374                 if (r == 4) {
375                         /*
376                          * Check action first
377                          */
378                         if (!strcmp(act, "wait"))
379                                 rw = DDIR_WAIT;
380                         else if (!strcmp(act, "read"))
381                                 rw = DDIR_READ;
382                         else if (!strcmp(act, "write"))
383                                 rw = DDIR_WRITE;
384                         else if (!strcmp(act, "sync"))
385                                 rw = DDIR_SYNC;
386                         else if (!strcmp(act, "datasync"))
387                                 rw = DDIR_DATASYNC;
388                         else if (!strcmp(act, "trim"))
389                                 rw = DDIR_TRIM;
390                         else {
391                                 log_err("fio: bad iolog file action: %s\n",
392                                                                         act);
393                                 continue;
394                         }
395                         fileno = get_fileno(td, fname);
396                 } else if (r == 2) {
397                         rw = DDIR_INVAL;
398                         if (!strcmp(act, "add")) {
399                                 fileno = add_file(td, fname, 0, 1);
400                                 file_action = FIO_LOG_ADD_FILE;
401                                 continue;
402                         } else if (!strcmp(act, "open")) {
403                                 fileno = get_fileno(td, fname);
404                                 file_action = FIO_LOG_OPEN_FILE;
405                         } else if (!strcmp(act, "close")) {
406                                 fileno = get_fileno(td, fname);
407                                 file_action = FIO_LOG_CLOSE_FILE;
408                         } else {
409                                 log_err("fio: bad iolog file action: %s\n",
410                                                                         act);
411                                 continue;
412                         }
413                 } else {
414                         log_err("bad iolog2: %s", p);
415                         continue;
416                 }
417
418                 if (rw == DDIR_READ)
419                         reads++;
420                 else if (rw == DDIR_WRITE) {
421                         /*
422                          * Don't add a write for ro mode
423                          */
424                         if (read_only)
425                                 continue;
426                         writes++;
427                 } else if (rw == DDIR_WAIT) {
428                         waits++;
429                 } else if (rw == DDIR_INVAL) {
430                 } else if (!ddir_sync(rw)) {
431                         log_err("bad ddir: %d\n", rw);
432                         continue;
433                 }
434
435                 /*
436                  * Make note of file
437                  */
438                 ipo = malloc(sizeof(*ipo));
439                 init_ipo(ipo);
440                 ipo->ddir = rw;
441                 if (rw == DDIR_WAIT) {
442                         ipo->delay = offset;
443                 } else {
444                         ipo->offset = offset;
445                         ipo->len = bytes;
446                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw])
447                                 td->o.max_bs[rw] = bytes;
448                         ipo->fileno = fileno;
449                         ipo->file_action = file_action;
450                         td->o.size += bytes;
451                 }
452
453                 queue_io_piece(td, ipo);
454         }
455
456         free(str);
457         free(act);
458         free(rfname);
459
460         if (writes && read_only) {
461                 log_err("fio: <%s> skips replay of %d writes due to"
462                         " read-only\n", td->o.name, writes);
463                 writes = 0;
464         }
465
466         if (!reads && !writes && !waits)
467                 return 1;
468         else if (reads && !writes)
469                 td->o.td_ddir = TD_DDIR_READ;
470         else if (!reads && writes)
471                 td->o.td_ddir = TD_DDIR_WRITE;
472         else
473                 td->o.td_ddir = TD_DDIR_RW;
474
475         return 0;
476 }
477
478 /*
479  * open iolog, check version, and call appropriate parser
480  */
481 static int init_iolog_read(struct thread_data *td)
482 {
483         char buffer[256], *p;
484         FILE *f;
485         int ret;
486
487         f = fopen(td->o.read_iolog_file, "r");
488         if (!f) {
489                 perror("fopen read iolog");
490                 return 1;
491         }
492
493         p = fgets(buffer, sizeof(buffer), f);
494         if (!p) {
495                 td_verror(td, errno, "iolog read");
496                 log_err("fio: unable to read iolog\n");
497                 fclose(f);
498                 return 1;
499         }
500
501         /*
502          * version 2 of the iolog stores a specific string as the
503          * first line, check for that
504          */
505         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
506                 ret = read_iolog2(td, f);
507         else {
508                 log_err("fio: iolog version 1 is no longer supported\n");
509                 ret = 1;
510         }
511
512         fclose(f);
513         return ret;
514 }
515
516 /*
517  * Set up a log for storing io patterns.
518  */
519 static int init_iolog_write(struct thread_data *td)
520 {
521         struct fio_file *ff;
522         FILE *f;
523         unsigned int i;
524
525         f = fopen(td->o.write_iolog_file, "a");
526         if (!f) {
527                 perror("fopen write iolog");
528                 return 1;
529         }
530
531         /*
532          * That's it for writing, setup a log buffer and we're done.
533           */
534         td->iolog_f = f;
535         td->iolog_buf = malloc(8192);
536         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
537
538         /*
539          * write our version line
540          */
541         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
542                 perror("iolog init\n");
543                 return 1;
544         }
545
546         /*
547          * add all known files
548          */
549         for_each_file(td, ff, i)
550                 log_file(td, ff, FIO_LOG_ADD_FILE);
551
552         return 0;
553 }
554
555 int init_iolog(struct thread_data *td)
556 {
557         int ret = 0;
558
559         if (td->o.read_iolog_file) {
560                 int need_swap;
561
562                 /*
563                  * Check if it's a blktrace file and load that if possible.
564                  * Otherwise assume it's a normal log file and load that.
565                  */
566                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
567                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
568                 else
569                         ret = init_iolog_read(td);
570         } else if (td->o.write_iolog_file)
571                 ret = init_iolog_write(td);
572
573         if (ret)
574                 td_verror(td, EINVAL, "failed initializing iolog");
575
576         return ret;
577 }
578
579 void setup_log(struct io_log **log, struct log_params *p,
580                const char *filename)
581 {
582         struct io_log *l;
583         int i;
584         struct io_u_plat_entry *entry;
585         struct flist_head *list;
586
587         l = scalloc(1, sizeof(*l));
588         INIT_FLIST_HEAD(&l->io_logs);
589         l->log_type = p->log_type;
590         l->log_offset = p->log_offset;
591         l->log_gz = p->log_gz;
592         l->log_gz_store = p->log_gz_store;
593         l->avg_msec = p->avg_msec;
594         l->hist_msec = p->hist_msec;
595         l->hist_coarseness = p->hist_coarseness;
596         l->filename = strdup(filename);
597         l->td = p->td;
598
599         /* Initialize histogram lists for each r/w direction,
600          * with initial io_u_plat of all zeros:
601          */
602         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
603                 list = &l->hist_window[i].list;
604                 INIT_FLIST_HEAD(list);
605                 entry = calloc(1, sizeof(struct io_u_plat_entry));
606                 flist_add(&entry->list, list);
607         }
608
609         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
610                 struct io_logs *p;
611
612                 p = calloc(1, sizeof(*l->pending));
613                 p->max_samples = DEF_LOG_ENTRIES;
614                 p->log = calloc(p->max_samples, log_entry_sz(l));
615                 l->pending = p;
616         }
617
618         if (l->log_offset)
619                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
620
621         INIT_FLIST_HEAD(&l->chunk_list);
622
623         if (l->log_gz && !p->td)
624                 l->log_gz = 0;
625         else if (l->log_gz || l->log_gz_store) {
626                 mutex_init_pshared(&l->chunk_lock);
627                 p->td->flags |= TD_F_COMPRESS_LOG;
628         }
629
630         *log = l;
631 }
632
633 #ifdef CONFIG_SETVBUF
634 static void *set_file_buffer(FILE *f)
635 {
636         size_t size = 1048576;
637         void *buf;
638
639         buf = malloc(size);
640         setvbuf(f, buf, _IOFBF, size);
641         return buf;
642 }
643
644 static void clear_file_buffer(void *buf)
645 {
646         free(buf);
647 }
648 #else
649 static void *set_file_buffer(FILE *f)
650 {
651         return NULL;
652 }
653
654 static void clear_file_buffer(void *buf)
655 {
656 }
657 #endif
658
659 void free_log(struct io_log *log)
660 {
661         while (!flist_empty(&log->io_logs)) {
662                 struct io_logs *cur_log;
663
664                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
665                 flist_del_init(&cur_log->list);
666                 free(cur_log->log);
667                 sfree(cur_log);
668         }
669
670         if (log->pending) {
671                 free(log->pending->log);
672                 free(log->pending);
673                 log->pending = NULL;
674         }
675
676         free(log->pending);
677         free(log->filename);
678         sfree(log);
679 }
680
681 inline unsigned long hist_sum(int j, int stride, unsigned int *io_u_plat,
682                 unsigned int *io_u_plat_last)
683 {
684         unsigned long sum;
685         int k;
686
687         if (io_u_plat_last) {
688                 for (k = sum = 0; k < stride; k++)
689                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
690         } else {
691                 for (k = sum = 0; k < stride; k++)
692                         sum += io_u_plat[j + k];
693         }
694
695         return sum;
696 }
697
698 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
699                                uint64_t sample_size)
700 {
701         struct io_sample *s;
702         int log_offset;
703         uint64_t i, j, nr_samples;
704         struct io_u_plat_entry *entry, *entry_before;
705         unsigned int *io_u_plat;
706         unsigned int *io_u_plat_before;
707
708         int stride = 1 << hist_coarseness;
709         
710         if (!sample_size)
711                 return;
712
713         s = __get_sample(samples, 0, 0);
714         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
715
716         nr_samples = sample_size / __log_entry_sz(log_offset);
717
718         for (i = 0; i < nr_samples; i++) {
719                 s = __get_sample(samples, log_offset, i);
720
721                 entry = (struct io_u_plat_entry *) (uintptr_t) s->val;
722                 io_u_plat = entry->io_u_plat;
723
724                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
725                 io_u_plat_before = entry_before->io_u_plat;
726
727                 fprintf(f, "%lu, %u, %u, ", (unsigned long) s->time,
728                                                 io_sample_ddir(s), s->bs);
729                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
730                         fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat,
731                                                 io_u_plat_before));
732                 }
733                 fprintf(f, "%lu\n", (unsigned long)
734                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
735                                         io_u_plat_before));
736
737                 flist_del(&entry_before->list);
738                 free(entry_before);
739         }
740 }
741
742 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
743 {
744         struct io_sample *s;
745         int log_offset;
746         uint64_t i, nr_samples;
747
748         if (!sample_size)
749                 return;
750
751         s = __get_sample(samples, 0, 0);
752         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
753
754         nr_samples = sample_size / __log_entry_sz(log_offset);
755
756         for (i = 0; i < nr_samples; i++) {
757                 s = __get_sample(samples, log_offset, i);
758
759                 if (!log_offset) {
760                         fprintf(f, "%lu, %lu, %u, %u\n",
761                                         (unsigned long) s->time,
762                                         (unsigned long) s->val,
763                                         io_sample_ddir(s), s->bs);
764                 } else {
765                         struct io_sample_offset *so = (void *) s;
766
767                         fprintf(f, "%lu, %lu, %u, %u, %llu\n",
768                                         (unsigned long) s->time,
769                                         (unsigned long) s->val,
770                                         io_sample_ddir(s), s->bs,
771                                         (unsigned long long) so->offset);
772                 }
773         }
774 }
775
776 #ifdef CONFIG_ZLIB
777
778 struct iolog_flush_data {
779         struct workqueue_work work;
780         struct io_log *log;
781         void *samples;
782         uint32_t nr_samples;
783         bool free;
784 };
785
786 #define GZ_CHUNK        131072
787
788 static struct iolog_compress *get_new_chunk(unsigned int seq)
789 {
790         struct iolog_compress *c;
791
792         c = malloc(sizeof(*c));
793         INIT_FLIST_HEAD(&c->list);
794         c->buf = malloc(GZ_CHUNK);
795         c->len = 0;
796         c->seq = seq;
797         return c;
798 }
799
800 static void free_chunk(struct iolog_compress *ic)
801 {
802         free(ic->buf);
803         free(ic);
804 }
805
806 static int z_stream_init(z_stream *stream, int gz_hdr)
807 {
808         int wbits = 15;
809
810         memset(stream, 0, sizeof(*stream));
811         stream->zalloc = Z_NULL;
812         stream->zfree = Z_NULL;
813         stream->opaque = Z_NULL;
814         stream->next_in = Z_NULL;
815
816         /*
817          * zlib magic - add 32 for auto-detection of gz header or not,
818          * if we decide to store files in a gzip friendly format.
819          */
820         if (gz_hdr)
821                 wbits += 32;
822
823         if (inflateInit2(stream, wbits) != Z_OK)
824                 return 1;
825
826         return 0;
827 }
828
829 struct inflate_chunk_iter {
830         unsigned int seq;
831         int err;
832         void *buf;
833         size_t buf_size;
834         size_t buf_used;
835         size_t chunk_sz;
836 };
837
838 static void finish_chunk(z_stream *stream, FILE *f,
839                          struct inflate_chunk_iter *iter)
840 {
841         int ret;
842
843         ret = inflateEnd(stream);
844         if (ret != Z_OK)
845                 log_err("fio: failed to end log inflation seq %d (%d)\n",
846                                 iter->seq, ret);
847
848         flush_samples(f, iter->buf, iter->buf_used);
849         free(iter->buf);
850         iter->buf = NULL;
851         iter->buf_size = iter->buf_used = 0;
852 }
853
854 /*
855  * Iterative chunk inflation. Handles cases where we cross into a new
856  * sequence, doing flush finish of previous chunk if needed.
857  */
858 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
859                             z_stream *stream, struct inflate_chunk_iter *iter)
860 {
861         size_t ret;
862
863         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
864                                 (unsigned long) ic->len, ic->seq);
865
866         if (ic->seq != iter->seq) {
867                 if (iter->seq)
868                         finish_chunk(stream, f, iter);
869
870                 z_stream_init(stream, gz_hdr);
871                 iter->seq = ic->seq;
872         }
873
874         stream->avail_in = ic->len;
875         stream->next_in = ic->buf;
876
877         if (!iter->buf_size) {
878                 iter->buf_size = iter->chunk_sz;
879                 iter->buf = malloc(iter->buf_size);
880         }
881
882         while (stream->avail_in) {
883                 size_t this_out = iter->buf_size - iter->buf_used;
884                 int err;
885
886                 stream->avail_out = this_out;
887                 stream->next_out = iter->buf + iter->buf_used;
888
889                 err = inflate(stream, Z_NO_FLUSH);
890                 if (err < 0) {
891                         log_err("fio: failed inflating log: %d\n", err);
892                         iter->err = err;
893                         break;
894                 }
895
896                 iter->buf_used += this_out - stream->avail_out;
897
898                 if (!stream->avail_out) {
899                         iter->buf_size += iter->chunk_sz;
900                         iter->buf = realloc(iter->buf, iter->buf_size);
901                         continue;
902                 }
903
904                 if (err == Z_STREAM_END)
905                         break;
906         }
907
908         ret = (void *) stream->next_in - ic->buf;
909
910         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
911
912         return ret;
913 }
914
915 /*
916  * Inflate stored compressed chunks, or write them directly to the log
917  * file if so instructed.
918  */
919 static int inflate_gz_chunks(struct io_log *log, FILE *f)
920 {
921         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
922         z_stream stream;
923
924         while (!flist_empty(&log->chunk_list)) {
925                 struct iolog_compress *ic;
926
927                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
928                 flist_del(&ic->list);
929
930                 if (log->log_gz_store) {
931                         size_t ret;
932
933                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
934                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
935
936                         ret = fwrite(ic->buf, ic->len, 1, f);
937                         if (ret != 1 || ferror(f)) {
938                                 iter.err = errno;
939                                 log_err("fio: error writing compressed log\n");
940                         }
941                 } else
942                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
943
944                 free_chunk(ic);
945         }
946
947         if (iter.seq) {
948                 finish_chunk(&stream, f, &iter);
949                 free(iter.buf);
950         }
951
952         return iter.err;
953 }
954
955 /*
956  * Open compressed log file and decompress the stored chunks and
957  * write them to stdout. The chunks are stored sequentially in the
958  * file, so we iterate over them and do them one-by-one.
959  */
960 int iolog_file_inflate(const char *file)
961 {
962         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
963         struct iolog_compress ic;
964         z_stream stream;
965         struct stat sb;
966         ssize_t ret;
967         size_t total;
968         void *buf;
969         FILE *f;
970
971         f = fopen(file, "r");
972         if (!f) {
973                 perror("fopen");
974                 return 1;
975         }
976
977         if (stat(file, &sb) < 0) {
978                 fclose(f);
979                 perror("stat");
980                 return 1;
981         }
982
983         ic.buf = buf = malloc(sb.st_size);
984         ic.len = sb.st_size;
985         ic.seq = 1;
986
987         ret = fread(ic.buf, ic.len, 1, f);
988         if (ret < 0) {
989                 perror("fread");
990                 fclose(f);
991                 free(buf);
992                 return 1;
993         } else if (ret != 1) {
994                 log_err("fio: short read on reading log\n");
995                 fclose(f);
996                 free(buf);
997                 return 1;
998         }
999
1000         fclose(f);
1001
1002         /*
1003          * Each chunk will return Z_STREAM_END. We don't know how many
1004          * chunks are in the file, so we just keep looping and incrementing
1005          * the sequence number until we have consumed the whole compressed
1006          * file.
1007          */
1008         total = ic.len;
1009         do {
1010                 size_t iret;
1011
1012                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1013                 total -= iret;
1014                 if (!total)
1015                         break;
1016                 if (iter.err)
1017                         break;
1018
1019                 ic.seq++;
1020                 ic.len -= iret;
1021                 ic.buf += iret;
1022         } while (1);
1023
1024         if (iter.seq) {
1025                 finish_chunk(&stream, stdout, &iter);
1026                 free(iter.buf);
1027         }
1028
1029         free(buf);
1030         return iter.err;
1031 }
1032
1033 #else
1034
1035 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1036 {
1037         return 0;
1038 }
1039
1040 int iolog_file_inflate(const char *file)
1041 {
1042         log_err("fio: log inflation not possible without zlib\n");
1043         return 1;
1044 }
1045
1046 #endif
1047
1048 void flush_log(struct io_log *log, bool do_append)
1049 {
1050         void *buf;
1051         FILE *f;
1052
1053         if (!do_append)
1054                 f = fopen(log->filename, "w");
1055         else
1056                 f = fopen(log->filename, "a");
1057         if (!f) {
1058                 perror("fopen log");
1059                 return;
1060         }
1061
1062         buf = set_file_buffer(f);
1063
1064         inflate_gz_chunks(log, f);
1065
1066         while (!flist_empty(&log->io_logs)) {
1067                 struct io_logs *cur_log;
1068
1069                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1070                 flist_del_init(&cur_log->list);
1071                 
1072                 if (log->td && log == log->td->clat_hist_log)
1073                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1074                                            log_sample_sz(log, cur_log));
1075                 else
1076                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1077                 
1078                 sfree(cur_log);
1079         }
1080
1081         fclose(f);
1082         clear_file_buffer(buf);
1083 }
1084
1085 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1086 {
1087         if (td->flags & TD_F_COMPRESS_LOG)
1088                 iolog_flush(log);
1089
1090         if (trylock) {
1091                 if (fio_trylock_file(log->filename))
1092                         return 1;
1093         } else
1094                 fio_lock_file(log->filename);
1095
1096         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1097                 fio_send_iolog(td, log, log->filename);
1098         else
1099                 flush_log(log, !td->o.per_job_logs);
1100
1101         fio_unlock_file(log->filename);
1102         free_log(log);
1103         return 0;
1104 }
1105
1106 size_t log_chunk_sizes(struct io_log *log)
1107 {
1108         struct flist_head *entry;
1109         size_t ret;
1110
1111         if (flist_empty(&log->chunk_list))
1112                 return 0;
1113
1114         ret = 0;
1115         pthread_mutex_lock(&log->chunk_lock);
1116         flist_for_each(entry, &log->chunk_list) {
1117                 struct iolog_compress *c;
1118
1119                 c = flist_entry(entry, struct iolog_compress, list);
1120                 ret += c->len;
1121         }
1122         pthread_mutex_unlock(&log->chunk_lock);
1123         return ret;
1124 }
1125
1126 #ifdef CONFIG_ZLIB
1127
1128 static int gz_work(struct iolog_flush_data *data)
1129 {
1130         struct iolog_compress *c = NULL;
1131         struct flist_head list;
1132         unsigned int seq;
1133         z_stream stream;
1134         size_t total = 0;
1135         int ret;
1136
1137         INIT_FLIST_HEAD(&list);
1138
1139         memset(&stream, 0, sizeof(stream));
1140         stream.zalloc = Z_NULL;
1141         stream.zfree = Z_NULL;
1142         stream.opaque = Z_NULL;
1143
1144         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1145         if (ret != Z_OK) {
1146                 log_err("fio: failed to init gz stream\n");
1147                 goto err;
1148         }
1149
1150         seq = ++data->log->chunk_seq;
1151
1152         stream.next_in = (void *) data->samples;
1153         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1154
1155         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1156                                 (unsigned long) stream.avail_in, seq,
1157                                 data->log->filename);
1158         do {
1159                 if (c)
1160                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1161                                 (unsigned long) c->len);
1162                 c = get_new_chunk(seq);
1163                 stream.avail_out = GZ_CHUNK;
1164                 stream.next_out = c->buf;
1165                 ret = deflate(&stream, Z_NO_FLUSH);
1166                 if (ret < 0) {
1167                         log_err("fio: deflate log (%d)\n", ret);
1168                         free_chunk(c);
1169                         goto err;
1170                 }
1171
1172                 c->len = GZ_CHUNK - stream.avail_out;
1173                 flist_add_tail(&c->list, &list);
1174                 total += c->len;
1175         } while (stream.avail_in);
1176
1177         stream.next_out = c->buf + c->len;
1178         stream.avail_out = GZ_CHUNK - c->len;
1179
1180         ret = deflate(&stream, Z_FINISH);
1181         if (ret < 0) {
1182                 /*
1183                  * Z_BUF_ERROR is special, it just means we need more
1184                  * output space. We'll handle that below. Treat any other
1185                  * error as fatal.
1186                  */
1187                 if (ret != Z_BUF_ERROR) {
1188                         log_err("fio: deflate log (%d)\n", ret);
1189                         flist_del(&c->list);
1190                         free_chunk(c);
1191                         goto err;
1192                 }
1193         }
1194
1195         total -= c->len;
1196         c->len = GZ_CHUNK - stream.avail_out;
1197         total += c->len;
1198         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1199
1200         if (ret != Z_STREAM_END) {
1201                 do {
1202                         c = get_new_chunk(seq);
1203                         stream.avail_out = GZ_CHUNK;
1204                         stream.next_out = c->buf;
1205                         ret = deflate(&stream, Z_FINISH);
1206                         c->len = GZ_CHUNK - stream.avail_out;
1207                         total += c->len;
1208                         flist_add_tail(&c->list, &list);
1209                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1210                                 (unsigned long) c->len);
1211                 } while (ret != Z_STREAM_END);
1212         }
1213
1214         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1215
1216         ret = deflateEnd(&stream);
1217         if (ret != Z_OK)
1218                 log_err("fio: deflateEnd %d\n", ret);
1219
1220         free(data->samples);
1221
1222         if (!flist_empty(&list)) {
1223                 pthread_mutex_lock(&data->log->chunk_lock);
1224                 flist_splice_tail(&list, &data->log->chunk_list);
1225                 pthread_mutex_unlock(&data->log->chunk_lock);
1226         }
1227
1228         ret = 0;
1229 done:
1230         if (data->free)
1231                 free(data);
1232         return ret;
1233 err:
1234         while (!flist_empty(&list)) {
1235                 c = flist_first_entry(list.next, struct iolog_compress, list);
1236                 flist_del(&c->list);
1237                 free_chunk(c);
1238         }
1239         ret = 1;
1240         goto done;
1241 }
1242
1243 /*
1244  * Invoked from our compress helper thread, when logging would have exceeded
1245  * the specified memory limitation. Compresses the previously stored
1246  * entries.
1247  */
1248 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1249 {
1250         return gz_work(container_of(work, struct iolog_flush_data, work));
1251 }
1252
1253 static int gz_init_worker(struct submit_worker *sw)
1254 {
1255         struct thread_data *td = sw->wq->td;
1256
1257         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1258                 return 0;
1259
1260         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1261                 log_err("gz: failed to set CPU affinity\n");
1262                 return 1;
1263         }
1264
1265         return 0;
1266 }
1267
1268 static struct workqueue_ops log_compress_wq_ops = {
1269         .fn             = gz_work_async,
1270         .init_worker_fn = gz_init_worker,
1271         .nice           = 1,
1272 };
1273
1274 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1275 {
1276         if (!(td->flags & TD_F_COMPRESS_LOG))
1277                 return 0;
1278
1279         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1280         return 0;
1281 }
1282
1283 void iolog_compress_exit(struct thread_data *td)
1284 {
1285         if (!(td->flags & TD_F_COMPRESS_LOG))
1286                 return;
1287
1288         workqueue_exit(&td->log_compress_wq);
1289 }
1290
1291 /*
1292  * Queue work item to compress the existing log entries. We reset the
1293  * current log to a small size, and reference the existing log in the
1294  * data that we queue for compression. Once compression has been done,
1295  * this old log is freed. If called with finish == true, will not return
1296  * until the log compression has completed, and will flush all previous
1297  * logs too
1298  */
1299 static int iolog_flush(struct io_log *log)
1300 {
1301         struct iolog_flush_data *data;
1302
1303         data = malloc(sizeof(*data));
1304         if (!data)
1305                 return 1;
1306
1307         data->log = log;
1308         data->free = false;
1309
1310         while (!flist_empty(&log->io_logs)) {
1311                 struct io_logs *cur_log;
1312
1313                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1314                 flist_del_init(&cur_log->list);
1315
1316                 data->samples = cur_log->log;
1317                 data->nr_samples = cur_log->nr_samples;
1318
1319                 sfree(cur_log);
1320
1321                 gz_work(data);
1322         }
1323
1324         free(data);
1325         return 0;
1326 }
1327
1328 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1329 {
1330         struct iolog_flush_data *data;
1331
1332         data = malloc(sizeof(*data));
1333         if (!data)
1334                 return 1;
1335
1336         data->log = log;
1337
1338         data->samples = cur_log->log;
1339         data->nr_samples = cur_log->nr_samples;
1340         data->free = true;
1341
1342         cur_log->nr_samples = cur_log->max_samples = 0;
1343         cur_log->log = NULL;
1344
1345         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1346         return 0;
1347 }
1348 #else
1349
1350 static int iolog_flush(struct io_log *log)
1351 {
1352         return 1;
1353 }
1354
1355 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1356 {
1357         return 1;
1358 }
1359
1360 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1361 {
1362         return 0;
1363 }
1364
1365 void iolog_compress_exit(struct thread_data *td)
1366 {
1367 }
1368
1369 #endif
1370
1371 struct io_logs *iolog_cur_log(struct io_log *log)
1372 {
1373         if (flist_empty(&log->io_logs))
1374                 return NULL;
1375
1376         return flist_last_entry(&log->io_logs, struct io_logs, list);
1377 }
1378
1379 uint64_t iolog_nr_samples(struct io_log *iolog)
1380 {
1381         struct flist_head *entry;
1382         uint64_t ret = 0;
1383
1384         flist_for_each(entry, &iolog->io_logs) {
1385                 struct io_logs *cur_log;
1386
1387                 cur_log = flist_entry(entry, struct io_logs, list);
1388                 ret += cur_log->nr_samples;
1389         }
1390
1391         return ret;
1392 }
1393
1394 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1395 {
1396         if (log)
1397                 return finish_log(td, log, try);
1398
1399         return 0;
1400 }
1401
1402 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1403 {
1404         int ret;
1405
1406         if (per_unit_log(td->iops_log) != unit_log)
1407                 return 0;
1408
1409         ret = __write_log(td, td->iops_log, try);
1410         if (!ret)
1411                 td->iops_log = NULL;
1412
1413         return ret;
1414 }
1415
1416 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1417 {
1418         int ret;
1419
1420         if (!unit_log)
1421                 return 0;
1422
1423         ret = __write_log(td, td->slat_log, try);
1424         if (!ret)
1425                 td->slat_log = NULL;
1426
1427         return ret;
1428 }
1429
1430 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1431 {
1432         int ret;
1433
1434         if (!unit_log)
1435                 return 0;
1436
1437         ret = __write_log(td, td->clat_log, try);
1438         if (!ret)
1439                 td->clat_log = NULL;
1440
1441         return ret;
1442 }
1443
1444 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1445 {
1446         int ret;
1447
1448         if (!unit_log)
1449                 return 0;
1450
1451         ret = __write_log(td, td->clat_hist_log, try);
1452         if (!ret)
1453                 td->clat_hist_log = NULL;
1454
1455         return ret;
1456 }
1457
1458 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1459 {
1460         int ret;
1461
1462         if (!unit_log)
1463                 return 0;
1464
1465         ret = __write_log(td, td->lat_log, try);
1466         if (!ret)
1467                 td->lat_log = NULL;
1468
1469         return ret;
1470 }
1471
1472 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1473 {
1474         int ret;
1475
1476         if (per_unit_log(td->bw_log) != unit_log)
1477                 return 0;
1478
1479         ret = __write_log(td, td->bw_log, try);
1480         if (!ret)
1481                 td->bw_log = NULL;
1482
1483         return ret;
1484 }
1485
1486 enum {
1487         BW_LOG_MASK     = 1,
1488         LAT_LOG_MASK    = 2,
1489         SLAT_LOG_MASK   = 4,
1490         CLAT_LOG_MASK   = 8,
1491         IOPS_LOG_MASK   = 16,
1492         CLAT_HIST_LOG_MASK = 32,
1493
1494         ALL_LOG_NR      = 6,
1495 };
1496
1497 struct log_type {
1498         unsigned int mask;
1499         int (*fn)(struct thread_data *, int, bool);
1500 };
1501
1502 static struct log_type log_types[] = {
1503         {
1504                 .mask   = BW_LOG_MASK,
1505                 .fn     = write_bandw_log,
1506         },
1507         {
1508                 .mask   = LAT_LOG_MASK,
1509                 .fn     = write_lat_log,
1510         },
1511         {
1512                 .mask   = SLAT_LOG_MASK,
1513                 .fn     = write_slat_log,
1514         },
1515         {
1516                 .mask   = CLAT_LOG_MASK,
1517                 .fn     = write_clat_log,
1518         },
1519         {
1520                 .mask   = IOPS_LOG_MASK,
1521                 .fn     = write_iops_log,
1522         },
1523         {
1524                 .mask   = CLAT_HIST_LOG_MASK,
1525                 .fn     = write_clat_hist_log,
1526         }
1527 };
1528
1529 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1530 {
1531         unsigned int log_mask = 0;
1532         unsigned int log_left = ALL_LOG_NR;
1533         int old_state, i;
1534
1535         old_state = td_bump_runstate(td, TD_FINISHING);
1536
1537         finalize_logs(td, unit_logs);
1538
1539         while (log_left) {
1540                 int prev_log_left = log_left;
1541
1542                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1543                         struct log_type *lt = &log_types[i];
1544                         int ret;
1545
1546                         if (!(log_mask & lt->mask)) {
1547                                 ret = lt->fn(td, log_left != 1, unit_logs);
1548                                 if (!ret) {
1549                                         log_left--;
1550                                         log_mask |= lt->mask;
1551                                 }
1552                         }
1553                 }
1554
1555                 if (prev_log_left == log_left)
1556                         usleep(5000);
1557         }
1558
1559         td_restore_runstate(td, old_state);
1560 }
1561
1562 void fio_writeout_logs(bool unit_logs)
1563 {
1564         struct thread_data *td;
1565         int i;
1566
1567         for_each_td(td, i)
1568                 td_writeout_logs(td, unit_logs);
1569 }