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