Fix occasional hangs on mutexes
[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         pthread_mutexattr_t mattr;
580
581         l = scalloc(1, sizeof(*l));
582         INIT_FLIST_HEAD(&l->io_logs);
583         l->log_type = p->log_type;
584         l->log_offset = p->log_offset;
585         l->log_gz = p->log_gz;
586         l->log_gz_store = p->log_gz_store;
587         l->avg_msec = p->avg_msec;
588         l->filename = strdup(filename);
589         l->td = p->td;
590
591         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
592                 struct io_logs *p;
593
594                 p = calloc(1, sizeof(*l->pending));
595                 p->max_samples = DEF_LOG_ENTRIES;
596                 p->log = calloc(p->max_samples, log_entry_sz(l));
597                 l->pending = p;
598         }
599
600         if (l->log_offset)
601                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
602
603         INIT_FLIST_HEAD(&l->chunk_list);
604
605         if (l->log_gz && !p->td)
606                 l->log_gz = 0;
607         else if (l->log_gz || l->log_gz_store) {
608                 pthread_mutexattr_init(&mattr);
609 #ifdef FIO_HAVE_PSHARED_MUTEX
610                 pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
611 #endif
612                 pthread_mutex_init(&l->chunk_lock, &mattr);
613                 p->td->flags |= TD_F_COMPRESS_LOG;
614         }
615
616         *log = l;
617 }
618
619 #ifdef CONFIG_SETVBUF
620 static void *set_file_buffer(FILE *f)
621 {
622         size_t size = 1048576;
623         void *buf;
624
625         buf = malloc(size);
626         setvbuf(f, buf, _IOFBF, size);
627         return buf;
628 }
629
630 static void clear_file_buffer(void *buf)
631 {
632         free(buf);
633 }
634 #else
635 static void *set_file_buffer(FILE *f)
636 {
637         return NULL;
638 }
639
640 static void clear_file_buffer(void *buf)
641 {
642 }
643 #endif
644
645 void free_log(struct io_log *log)
646 {
647         while (!flist_empty(&log->io_logs)) {
648                 struct io_logs *cur_log;
649
650                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
651                 flist_del_init(&cur_log->list);
652                 free(cur_log->log);
653         }
654
655         if (log->pending) {
656                 free(log->pending->log);
657                 free(log->pending);
658                 log->pending = NULL;
659         }
660
661         free(log->pending);
662         free(log->filename);
663         sfree(log);
664 }
665
666 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
667 {
668         struct io_sample *s;
669         int log_offset;
670         uint64_t i, nr_samples;
671
672         if (!sample_size)
673                 return;
674
675         s = __get_sample(samples, 0, 0);
676         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
677
678         nr_samples = sample_size / __log_entry_sz(log_offset);
679
680         for (i = 0; i < nr_samples; i++) {
681                 s = __get_sample(samples, log_offset, i);
682
683                 if (!log_offset) {
684                         fprintf(f, "%lu, %lu, %u, %u\n",
685                                         (unsigned long) s->time,
686                                         (unsigned long) s->val,
687                                         io_sample_ddir(s), s->bs);
688                 } else {
689                         struct io_sample_offset *so = (void *) s;
690
691                         fprintf(f, "%lu, %lu, %u, %u, %llu\n",
692                                         (unsigned long) s->time,
693                                         (unsigned long) s->val,
694                                         io_sample_ddir(s), s->bs,
695                                         (unsigned long long) so->offset);
696                 }
697         }
698 }
699
700 #ifdef CONFIG_ZLIB
701
702 struct iolog_flush_data {
703         struct workqueue_work work;
704         struct io_log *log;
705         void *samples;
706         uint32_t nr_samples;
707         bool free;
708 };
709
710 #define GZ_CHUNK        131072
711
712 static struct iolog_compress *get_new_chunk(unsigned int seq)
713 {
714         struct iolog_compress *c;
715
716         c = malloc(sizeof(*c));
717         INIT_FLIST_HEAD(&c->list);
718         c->buf = malloc(GZ_CHUNK);
719         c->len = 0;
720         c->seq = seq;
721         return c;
722 }
723
724 static void free_chunk(struct iolog_compress *ic)
725 {
726         free(ic->buf);
727         free(ic);
728 }
729
730 static int z_stream_init(z_stream *stream, int gz_hdr)
731 {
732         int wbits = 15;
733
734         memset(stream, 0, sizeof(*stream));
735         stream->zalloc = Z_NULL;
736         stream->zfree = Z_NULL;
737         stream->opaque = Z_NULL;
738         stream->next_in = Z_NULL;
739
740         /*
741          * zlib magic - add 32 for auto-detection of gz header or not,
742          * if we decide to store files in a gzip friendly format.
743          */
744         if (gz_hdr)
745                 wbits += 32;
746
747         if (inflateInit2(stream, wbits) != Z_OK)
748                 return 1;
749
750         return 0;
751 }
752
753 struct inflate_chunk_iter {
754         unsigned int seq;
755         int err;
756         void *buf;
757         size_t buf_size;
758         size_t buf_used;
759         size_t chunk_sz;
760 };
761
762 static void finish_chunk(z_stream *stream, FILE *f,
763                          struct inflate_chunk_iter *iter)
764 {
765         int ret;
766
767         ret = inflateEnd(stream);
768         if (ret != Z_OK)
769                 log_err("fio: failed to end log inflation seq %d (%d)\n",
770                                 iter->seq, ret);
771
772         flush_samples(f, iter->buf, iter->buf_used);
773         free(iter->buf);
774         iter->buf = NULL;
775         iter->buf_size = iter->buf_used = 0;
776 }
777
778 /*
779  * Iterative chunk inflation. Handles cases where we cross into a new
780  * sequence, doing flush finish of previous chunk if needed.
781  */
782 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
783                             z_stream *stream, struct inflate_chunk_iter *iter)
784 {
785         size_t ret;
786
787         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
788                                 (unsigned long) ic->len, ic->seq);
789
790         if (ic->seq != iter->seq) {
791                 if (iter->seq)
792                         finish_chunk(stream, f, iter);
793
794                 z_stream_init(stream, gz_hdr);
795                 iter->seq = ic->seq;
796         }
797
798         stream->avail_in = ic->len;
799         stream->next_in = ic->buf;
800
801         if (!iter->buf_size) {
802                 iter->buf_size = iter->chunk_sz;
803                 iter->buf = malloc(iter->buf_size);
804         }
805
806         while (stream->avail_in) {
807                 size_t this_out = iter->buf_size - iter->buf_used;
808                 int err;
809
810                 stream->avail_out = this_out;
811                 stream->next_out = iter->buf + iter->buf_used;
812
813                 err = inflate(stream, Z_NO_FLUSH);
814                 if (err < 0) {
815                         log_err("fio: failed inflating log: %d\n", err);
816                         iter->err = err;
817                         break;
818                 }
819
820                 iter->buf_used += this_out - stream->avail_out;
821
822                 if (!stream->avail_out) {
823                         iter->buf_size += iter->chunk_sz;
824                         iter->buf = realloc(iter->buf, iter->buf_size);
825                         continue;
826                 }
827
828                 if (err == Z_STREAM_END)
829                         break;
830         }
831
832         ret = (void *) stream->next_in - ic->buf;
833
834         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
835
836         return ret;
837 }
838
839 /*
840  * Inflate stored compressed chunks, or write them directly to the log
841  * file if so instructed.
842  */
843 static int inflate_gz_chunks(struct io_log *log, FILE *f)
844 {
845         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
846         z_stream stream;
847
848         while (!flist_empty(&log->chunk_list)) {
849                 struct iolog_compress *ic;
850
851                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
852                 flist_del(&ic->list);
853
854                 if (log->log_gz_store) {
855                         size_t ret;
856
857                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
858                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
859
860                         ret = fwrite(ic->buf, ic->len, 1, f);
861                         if (ret != 1 || ferror(f)) {
862                                 iter.err = errno;
863                                 log_err("fio: error writing compressed log\n");
864                         }
865                 } else
866                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
867
868                 free_chunk(ic);
869         }
870
871         if (iter.seq) {
872                 finish_chunk(&stream, f, &iter);
873                 free(iter.buf);
874         }
875
876         return iter.err;
877 }
878
879 /*
880  * Open compressed log file and decompress the stored chunks and
881  * write them to stdout. The chunks are stored sequentially in the
882  * file, so we iterate over them and do them one-by-one.
883  */
884 int iolog_file_inflate(const char *file)
885 {
886         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
887         struct iolog_compress ic;
888         z_stream stream;
889         struct stat sb;
890         ssize_t ret;
891         size_t total;
892         void *buf;
893         FILE *f;
894
895         f = fopen(file, "r");
896         if (!f) {
897                 perror("fopen");
898                 return 1;
899         }
900
901         if (stat(file, &sb) < 0) {
902                 fclose(f);
903                 perror("stat");
904                 return 1;
905         }
906
907         ic.buf = buf = malloc(sb.st_size);
908         ic.len = sb.st_size;
909         ic.seq = 1;
910
911         ret = fread(ic.buf, ic.len, 1, f);
912         if (ret < 0) {
913                 perror("fread");
914                 fclose(f);
915                 free(buf);
916                 return 1;
917         } else if (ret != 1) {
918                 log_err("fio: short read on reading log\n");
919                 fclose(f);
920                 free(buf);
921                 return 1;
922         }
923
924         fclose(f);
925
926         /*
927          * Each chunk will return Z_STREAM_END. We don't know how many
928          * chunks are in the file, so we just keep looping and incrementing
929          * the sequence number until we have consumed the whole compressed
930          * file.
931          */
932         total = ic.len;
933         do {
934                 size_t iret;
935
936                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
937                 total -= iret;
938                 if (!total)
939                         break;
940                 if (iter.err)
941                         break;
942
943                 ic.seq++;
944                 ic.len -= iret;
945                 ic.buf += iret;
946         } while (1);
947
948         if (iter.seq) {
949                 finish_chunk(&stream, stdout, &iter);
950                 free(iter.buf);
951         }
952
953         free(buf);
954         return iter.err;
955 }
956
957 #else
958
959 static int inflate_gz_chunks(struct io_log *log, FILE *f)
960 {
961         return 0;
962 }
963
964 int iolog_file_inflate(const char *file)
965 {
966         log_err("fio: log inflation not possible without zlib\n");
967         return 1;
968 }
969
970 #endif
971
972 void flush_log(struct io_log *log, int do_append)
973 {
974         void *buf;
975         FILE *f;
976
977         if (!do_append)
978                 f = fopen(log->filename, "w");
979         else
980                 f = fopen(log->filename, "a");
981         if (!f) {
982                 perror("fopen log");
983                 return;
984         }
985
986         buf = set_file_buffer(f);
987
988         inflate_gz_chunks(log, f);
989
990         while (!flist_empty(&log->io_logs)) {
991                 struct io_logs *cur_log;
992
993                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
994                 flist_del_init(&cur_log->list);
995                 flush_samples(f, cur_log->log, cur_log->nr_samples * log_entry_sz(log));
996         }
997
998         fclose(f);
999         clear_file_buffer(buf);
1000 }
1001
1002 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1003 {
1004         if (td->flags & TD_F_COMPRESS_LOG)
1005                 iolog_flush(log);
1006
1007         if (trylock) {
1008                 if (fio_trylock_file(log->filename))
1009                         return 1;
1010         } else
1011                 fio_lock_file(log->filename);
1012
1013         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1014                 fio_send_iolog(td, log, log->filename);
1015         else
1016                 flush_log(log, !td->o.per_job_logs);
1017
1018         fio_unlock_file(log->filename);
1019         free_log(log);
1020         return 0;
1021 }
1022
1023 size_t log_chunk_sizes(struct io_log *log)
1024 {
1025         struct flist_head *entry;
1026         size_t ret;
1027
1028         if (flist_empty(&log->chunk_list))
1029                 return 0;
1030
1031         ret = 0;
1032         pthread_mutex_lock(&log->chunk_lock);
1033         flist_for_each(entry, &log->chunk_list) {
1034                 struct iolog_compress *c;
1035
1036                 c = flist_entry(entry, struct iolog_compress, list);
1037                 ret += c->len;
1038         }
1039         pthread_mutex_unlock(&log->chunk_lock);
1040         return ret;
1041 }
1042
1043 #ifdef CONFIG_ZLIB
1044
1045 static int gz_work(struct iolog_flush_data *data)
1046 {
1047         struct iolog_compress *c = NULL;
1048         struct flist_head list;
1049         unsigned int seq;
1050         z_stream stream;
1051         size_t total = 0;
1052         int ret;
1053
1054         INIT_FLIST_HEAD(&list);
1055
1056         memset(&stream, 0, sizeof(stream));
1057         stream.zalloc = Z_NULL;
1058         stream.zfree = Z_NULL;
1059         stream.opaque = Z_NULL;
1060
1061         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1062         if (ret != Z_OK) {
1063                 log_err("fio: failed to init gz stream\n");
1064                 goto err;
1065         }
1066
1067         seq = ++data->log->chunk_seq;
1068
1069         stream.next_in = (void *) data->samples;
1070         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1071
1072         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1073                                 (unsigned long) stream.avail_in, seq,
1074                                 data->log->filename);
1075         do {
1076                 if (c)
1077                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1078                 c = get_new_chunk(seq);
1079                 stream.avail_out = GZ_CHUNK;
1080                 stream.next_out = c->buf;
1081                 ret = deflate(&stream, Z_NO_FLUSH);
1082                 if (ret < 0) {
1083                         log_err("fio: deflate log (%d)\n", ret);
1084                         free_chunk(c);
1085                         goto err;
1086                 }
1087
1088                 c->len = GZ_CHUNK - stream.avail_out;
1089                 flist_add_tail(&c->list, &list);
1090                 total += c->len;
1091         } while (stream.avail_in);
1092
1093         stream.next_out = c->buf + c->len;
1094         stream.avail_out = GZ_CHUNK - c->len;
1095
1096         ret = deflate(&stream, Z_FINISH);
1097         if (ret < 0) {
1098                 /*
1099                  * Z_BUF_ERROR is special, it just means we need more
1100                  * output space. We'll handle that below. Treat any other
1101                  * error as fatal.
1102                  */
1103                 if (ret != Z_BUF_ERROR) {
1104                         log_err("fio: deflate log (%d)\n", ret);
1105                         flist_del(&c->list);
1106                         free_chunk(c);
1107                         goto err;
1108                 }
1109         }
1110
1111         total -= c->len;
1112         c->len = GZ_CHUNK - stream.avail_out;
1113         total += c->len;
1114         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1115
1116         if (ret != Z_STREAM_END) {
1117                 do {
1118                         c = get_new_chunk(seq);
1119                         stream.avail_out = GZ_CHUNK;
1120                         stream.next_out = c->buf;
1121                         ret = deflate(&stream, Z_FINISH);
1122                         c->len = GZ_CHUNK - stream.avail_out;
1123                         total += c->len;
1124                         flist_add_tail(&c->list, &list);
1125                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1126                 } while (ret != Z_STREAM_END);
1127         }
1128
1129         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1130
1131         ret = deflateEnd(&stream);
1132         if (ret != Z_OK)
1133                 log_err("fio: deflateEnd %d\n", ret);
1134
1135         free(data->samples);
1136
1137         if (!flist_empty(&list)) {
1138                 pthread_mutex_lock(&data->log->chunk_lock);
1139                 flist_splice_tail(&list, &data->log->chunk_list);
1140                 pthread_mutex_unlock(&data->log->chunk_lock);
1141         }
1142
1143         ret = 0;
1144 done:
1145         if (data->free)
1146                 free(data);
1147         return ret;
1148 err:
1149         while (!flist_empty(&list)) {
1150                 c = flist_first_entry(list.next, struct iolog_compress, list);
1151                 flist_del(&c->list);
1152                 free_chunk(c);
1153         }
1154         ret = 1;
1155         goto done;
1156 }
1157
1158 /*
1159  * Invoked from our compress helper thread, when logging would have exceeded
1160  * the specified memory limitation. Compresses the previously stored
1161  * entries.
1162  */
1163 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1164 {
1165         return gz_work(container_of(work, struct iolog_flush_data, work));
1166 }
1167
1168 static int gz_init_worker(struct submit_worker *sw)
1169 {
1170         struct thread_data *td = sw->wq->td;
1171
1172         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1173                 return 0;
1174
1175         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1176                 log_err("gz: failed to set CPU affinity\n");
1177                 return 1;
1178         }
1179
1180         return 0;
1181 }
1182
1183 static struct workqueue_ops log_compress_wq_ops = {
1184         .fn             = gz_work_async,
1185         .init_worker_fn = gz_init_worker,
1186         .nice           = 1,
1187 };
1188
1189 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1190 {
1191         if (!(td->flags & TD_F_COMPRESS_LOG))
1192                 return 0;
1193
1194         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1195         return 0;
1196 }
1197
1198 void iolog_compress_exit(struct thread_data *td)
1199 {
1200         if (!(td->flags & TD_F_COMPRESS_LOG))
1201                 return;
1202
1203         workqueue_exit(&td->log_compress_wq);
1204 }
1205
1206 /*
1207  * Queue work item to compress the existing log entries. We reset the
1208  * current log to a small size, and reference the existing log in the
1209  * data that we queue for compression. Once compression has been done,
1210  * this old log is freed. If called with finish == true, will not return
1211  * until the log compression has completed, and will flush all previous
1212  * logs too
1213  */
1214 static int iolog_flush(struct io_log *log)
1215 {
1216         struct iolog_flush_data *data;
1217
1218         data = malloc(sizeof(*data));
1219         if (!data)
1220                 return 1;
1221
1222         data->log = log;
1223         data->free = false;
1224
1225         while (!flist_empty(&log->io_logs)) {
1226                 struct io_logs *cur_log;
1227
1228                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1229                 flist_del_init(&cur_log->list);
1230
1231                 data->samples = cur_log->log;
1232                 data->nr_samples = cur_log->nr_samples;
1233
1234                 cur_log->nr_samples = 0;
1235                 cur_log->max_samples = 0;
1236                 cur_log->log = NULL;
1237
1238                 gz_work(data);
1239         }
1240
1241         free(data);
1242         return 0;
1243 }
1244
1245 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1246 {
1247         struct iolog_flush_data *data;
1248
1249         data = malloc(sizeof(*data));
1250         if (!data)
1251                 return 1;
1252
1253         data->log = log;
1254
1255         data->samples = cur_log->log;
1256         data->nr_samples = cur_log->nr_samples;
1257         data->free = true;
1258
1259         cur_log->nr_samples = cur_log->max_samples = 0;
1260         cur_log->log = NULL;
1261
1262         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1263         return 0;
1264 }
1265 #else
1266
1267 static int iolog_flush(struct io_log *log)
1268 {
1269         return 1;
1270 }
1271
1272 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1273 {
1274         return 1;
1275 }
1276
1277 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1278 {
1279         return 0;
1280 }
1281
1282 void iolog_compress_exit(struct thread_data *td)
1283 {
1284 }
1285
1286 #endif
1287
1288 struct io_logs *iolog_cur_log(struct io_log *log)
1289 {
1290         if (flist_empty(&log->io_logs))
1291                 return NULL;
1292
1293         return flist_last_entry(&log->io_logs, struct io_logs, list);
1294 }
1295
1296 uint64_t iolog_nr_samples(struct io_log *iolog)
1297 {
1298         struct flist_head *entry;
1299         uint64_t ret = 0;
1300
1301         flist_for_each(entry, &iolog->io_logs) {
1302                 struct io_logs *cur_log;
1303
1304                 cur_log = flist_entry(entry, struct io_logs, list);
1305                 ret += cur_log->nr_samples;
1306         }
1307
1308         return ret;
1309 }
1310
1311 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1312 {
1313         if (log)
1314                 return finish_log(td, log, try);
1315
1316         return 0;
1317 }
1318
1319 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1320 {
1321         int ret;
1322
1323         if (per_unit_log(td->iops_log) != unit_log)
1324                 return 0;
1325
1326         ret = __write_log(td, td->iops_log, try);
1327         if (!ret)
1328                 td->iops_log = NULL;
1329
1330         return ret;
1331 }
1332
1333 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1334 {
1335         int ret;
1336
1337         if (!unit_log)
1338                 return 0;
1339
1340         ret = __write_log(td, td->slat_log, try);
1341         if (!ret)
1342                 td->slat_log = NULL;
1343
1344         return ret;
1345 }
1346
1347 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1348 {
1349         int ret;
1350
1351         if (!unit_log)
1352                 return 0;
1353
1354         ret = __write_log(td, td->clat_log, try);
1355         if (!ret)
1356                 td->clat_log = NULL;
1357
1358         return ret;
1359 }
1360
1361 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1362 {
1363         int ret;
1364
1365         if (!unit_log)
1366                 return 0;
1367
1368         ret = __write_log(td, td->lat_log, try);
1369         if (!ret)
1370                 td->lat_log = NULL;
1371
1372         return ret;
1373 }
1374
1375 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1376 {
1377         int ret;
1378
1379         if (per_unit_log(td->bw_log) != unit_log)
1380                 return 0;
1381
1382         ret = __write_log(td, td->bw_log, try);
1383         if (!ret)
1384                 td->bw_log = NULL;
1385
1386         return ret;
1387 }
1388
1389 enum {
1390         BW_LOG_MASK     = 1,
1391         LAT_LOG_MASK    = 2,
1392         SLAT_LOG_MASK   = 4,
1393         CLAT_LOG_MASK   = 8,
1394         IOPS_LOG_MASK   = 16,
1395
1396         ALL_LOG_NR      = 5,
1397 };
1398
1399 struct log_type {
1400         unsigned int mask;
1401         int (*fn)(struct thread_data *, int, bool);
1402 };
1403
1404 static struct log_type log_types[] = {
1405         {
1406                 .mask   = BW_LOG_MASK,
1407                 .fn     = write_bandw_log,
1408         },
1409         {
1410                 .mask   = LAT_LOG_MASK,
1411                 .fn     = write_lat_log,
1412         },
1413         {
1414                 .mask   = SLAT_LOG_MASK,
1415                 .fn     = write_slat_log,
1416         },
1417         {
1418                 .mask   = CLAT_LOG_MASK,
1419                 .fn     = write_clat_log,
1420         },
1421         {
1422                 .mask   = IOPS_LOG_MASK,
1423                 .fn     = write_iops_log,
1424         },
1425 };
1426
1427 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1428 {
1429         unsigned int log_mask = 0;
1430         unsigned int log_left = ALL_LOG_NR;
1431         int old_state, i;
1432
1433         old_state = td_bump_runstate(td, TD_FINISHING);
1434
1435         finalize_logs(td, unit_logs);
1436
1437         while (log_left) {
1438                 int prev_log_left = log_left;
1439
1440                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1441                         struct log_type *lt = &log_types[i];
1442                         int ret;
1443
1444                         if (!(log_mask & lt->mask)) {
1445                                 ret = lt->fn(td, log_left != 1, unit_logs);
1446                                 if (!ret) {
1447                                         log_left--;
1448                                         log_mask |= lt->mask;
1449                                 }
1450                         }
1451                 }
1452
1453                 if (prev_log_left == log_left)
1454                         usleep(5000);
1455         }
1456
1457         td_restore_runstate(td, old_state);
1458 }
1459
1460 void fio_writeout_logs(bool unit_logs)
1461 {
1462         struct thread_data *td;
1463         int i;
1464
1465         for_each_td(td, i)
1466                 td_writeout_logs(td, unit_logs);
1467 }