2 * Code related to writing an iolog of what a thread is doing, and to
3 * later read that back and replay
23 static int iolog_flush(struct io_log *log);
25 static const char iolog_ver2[] = "fio version 2 iolog";
27 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
29 flist_add_tail(&ipo->list, &td->io_log_list);
30 td->total_io_size += ipo->len;
33 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
35 if (!td->o.write_iolog_file)
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);
43 void log_file(struct thread_data *td, struct fio_file *f,
44 enum file_log_act what)
46 const char *act[] = { "add", "open", "close" };
50 if (!td->o.write_iolog_file)
55 * this happens on the pre-open/close done before the job starts
60 fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
63 static void iolog_delay(struct thread_data *td, unsigned long delay)
65 uint64_t usec = utime_since_now(&td->last_issue);
69 if (delay < td->time_offset) {
74 delay -= td->time_offset;
80 fio_gettime(&tv, NULL);
81 while (delay && !td->terminate) {
83 if (this_delay > 500000)
86 usec_sleep(td, this_delay);
90 usec = utime_since_now(&tv);
92 td->time_offset = usec - delay;
97 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
105 if (ipo->ddir != DDIR_INVAL)
108 f = td->files[ipo->fileno];
110 switch (ipo->file_action) {
111 case FIO_LOG_OPEN_FILE:
112 ret = td_io_open_file(td, f);
115 td_verror(td, ret, "iolog open file");
117 case FIO_LOG_CLOSE_FILE:
118 td_io_close_file(td, f);
120 case FIO_LOG_UNLINK_FILE:
121 td_io_unlink_file(td, f);
124 log_err("fio: bad file action %d\n", ipo->file_action);
131 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
133 struct io_piece *ipo;
134 unsigned long elapsed;
136 while (!flist_empty(&td->io_log_list)) {
139 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
140 flist_del(&ipo->list);
141 remove_trim_entry(td, ipo);
143 ret = ipo_special(td, ipo);
147 } else if (ret > 0) {
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);
161 iolog_delay(td, ipo->delay);
163 elapsed = mtime_since_genesis();
164 if (ipo->delay > elapsed)
165 usec_sleep(td, (ipo->delay - elapsed) * 1000);
170 if (io_u->ddir != DDIR_WAIT)
178 void prune_io_piece_log(struct thread_data *td)
180 struct io_piece *ipo;
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);
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);
201 * log a successful write, so we can unwind the log for verify
203 void log_io_piece(struct thread_data *td, struct io_u *io_u)
205 struct rb_node **p, *parent;
206 struct io_piece *ipo, *__ipo;
208 ipo = malloc(sizeof(struct io_piece));
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;
218 if (io_u_should_trim(td, io_u)) {
219 flist_add_tail(&ipo->trim_list, &td->trim_list);
224 * We don't need to sort the entries, if:
226 * Sequential writes, or
227 * Random writes that lay out the file as it goes along
229 * For both these cases, just reading back data in the order we
230 * wrote it out is the fastest.
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
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;
246 RB_CLEAR_NODE(&ipo->rb_node);
249 * Sort the entry into the verification list
252 p = &td->io_hist_tree.rb_node;
258 __ipo = rb_entry(parent, struct io_piece, rb_node);
259 if (ipo->file < __ipo->file)
261 else if (ipo->file > __ipo->file)
263 else if (ipo->offset < __ipo->offset) {
265 overlap = ipo->offset + ipo->len > __ipo->offset;
267 else if (ipo->offset > __ipo->offset) {
269 overlap = __ipo->offset + __ipo->len > ipo->offset;
275 dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu",
276 __ipo->offset, __ipo->len,
277 ipo->offset, ipo->len);
279 rb_erase(parent, &td->io_hist_tree);
280 remove_trim_entry(td, __ipo);
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;
292 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
294 struct io_piece *ipo = io_u->ipo;
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);
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);
321 void trim_io_piece(struct thread_data *td, const struct io_u *io_u)
323 struct io_piece *ipo = io_u->ipo;
328 ipo->len = io_u->xfer_buflen - io_u->resid;
331 void write_iolog_close(struct thread_data *td)
337 td->iolog_buf = NULL;
341 * Read version 2 iolog data. It is enhanced to include per-file logging,
344 static int read_iolog2(struct thread_data *td, FILE *f)
346 unsigned long long offset;
348 int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
353 free_release_files(td);
356 * Read in the read iolog and store it, reuse the infrastructure
357 * for doing verifications.
360 fname = malloc(256+16);
361 act = malloc(256+16);
363 reads = writes = waits = 0;
364 while ((p = fgets(str, 4096, f)) != NULL) {
365 struct io_piece *ipo;
368 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
374 if (!strcmp(act, "wait"))
376 else if (!strcmp(act, "read"))
378 else if (!strcmp(act, "write"))
380 else if (!strcmp(act, "sync"))
382 else if (!strcmp(act, "datasync"))
384 else if (!strcmp(act, "trim"))
387 log_err("fio: bad iolog file action: %s\n",
391 fileno = get_fileno(td, fname);
394 if (!strcmp(act, "add")) {
395 fileno = add_file(td, fname, 0, 1);
396 file_action = FIO_LOG_ADD_FILE;
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;
405 log_err("fio: bad iolog file action: %s\n",
410 log_err("bad iolog2: %s", p);
416 else if (rw == DDIR_WRITE) {
418 * Don't add a write for ro mode
423 } else if (rw == DDIR_WAIT) {
425 } else if (rw == DDIR_INVAL) {
426 } else if (!ddir_sync(rw)) {
427 log_err("bad ddir: %d\n", rw);
434 ipo = malloc(sizeof(*ipo));
437 if (rw == DDIR_WAIT) {
440 ipo->offset = offset;
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;
449 queue_io_piece(td, ipo);
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);
462 if (!reads && !writes && !waits)
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;
469 td->o.td_ddir = TD_DDIR_RW;
475 * open iolog, check version, and call appropriate parser
477 static int init_iolog_read(struct thread_data *td)
479 char buffer[256], *p;
483 f = fopen(td->o.read_iolog_file, "r");
485 perror("fopen read iolog");
489 p = fgets(buffer, sizeof(buffer), f);
491 td_verror(td, errno, "iolog read");
492 log_err("fio: unable to read iolog\n");
498 * version 2 of the iolog stores a specific string as the
499 * first line, check for that
501 if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
502 ret = read_iolog2(td, f);
504 log_err("fio: iolog version 1 is no longer supported\n");
513 * Set up a log for storing io patterns.
515 static int init_iolog_write(struct thread_data *td)
521 f = fopen(td->o.write_iolog_file, "a");
523 perror("fopen write iolog");
528 * That's it for writing, setup a log buffer and we're done.
531 td->iolog_buf = malloc(8192);
532 setvbuf(f, td->iolog_buf, _IOFBF, 8192);
535 * write our version line
537 if (fprintf(f, "%s\n", iolog_ver2) < 0) {
538 perror("iolog init\n");
543 * add all known files
545 for_each_file(td, ff, i)
546 log_file(td, ff, FIO_LOG_ADD_FILE);
551 int init_iolog(struct thread_data *td)
555 if (td->o.read_iolog_file) {
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.
562 if (is_blktrace(td->o.read_iolog_file, &need_swap))
563 ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
565 ret = init_iolog_read(td);
566 } else if (td->o.write_iolog_file)
567 ret = init_iolog_write(td);
570 td_verror(td, EINVAL, "failed initializing iolog");
575 void setup_log(struct io_log **log, struct log_params *p,
576 const char *filename)
580 struct io_u_plat_entry *entry;
581 struct flist_head *list;
583 l = scalloc(1, sizeof(*l));
584 INIT_FLIST_HEAD(&l->io_logs);
585 l->log_type = p->log_type;
586 l->log_offset = p->log_offset;
587 l->log_gz = p->log_gz;
588 l->log_gz_store = p->log_gz_store;
589 l->avg_msec = p->avg_msec;
590 l->hist_msec = p->hist_msec;
591 l->hist_coarseness = p->hist_coarseness;
592 l->filename = strdup(filename);
595 /* Initialize histogram lists for each r/w direction,
596 * with initial io_u_plat of all zeros:
598 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
599 list = &l->hist_window[i].list.list;
600 INIT_FLIST_HEAD(list);
601 entry = calloc(1, sizeof(struct io_u_plat_entry));
602 flist_add(&entry->list, list);
605 if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
608 p = calloc(1, sizeof(*l->pending));
609 p->max_samples = DEF_LOG_ENTRIES;
610 p->log = calloc(p->max_samples, log_entry_sz(l));
615 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
617 INIT_FLIST_HEAD(&l->chunk_list);
619 if (l->log_gz && !p->td)
621 else if (l->log_gz || l->log_gz_store) {
622 mutex_init_pshared(&l->chunk_lock);
623 p->td->flags |= TD_F_COMPRESS_LOG;
629 #ifdef CONFIG_SETVBUF
630 static void *set_file_buffer(FILE *f)
632 size_t size = 1048576;
636 setvbuf(f, buf, _IOFBF, size);
640 static void clear_file_buffer(void *buf)
645 static void *set_file_buffer(FILE *f)
650 static void clear_file_buffer(void *buf)
655 void free_log(struct io_log *log)
657 while (!flist_empty(&log->io_logs)) {
658 struct io_logs *cur_log;
660 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
661 flist_del_init(&cur_log->list);
667 free(log->pending->log);
677 static inline unsigned long hist_sum(int j, int stride, unsigned int *io_u_plat,
678 unsigned int *io_u_plat_last)
683 for (k = sum = 0; k < stride; k++)
684 sum += io_u_plat[j + k] - io_u_plat_last[j + k];
689 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
690 uint64_t sample_size)
694 uint64_t i, j, nr_samples;
695 struct io_u_plat_entry *entry, *entry_before;
696 unsigned int *io_u_plat;
697 unsigned int *io_u_plat_before;
699 int stride = 1 << hist_coarseness;
704 s = __get_sample(samples, 0, 0);
705 log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
707 nr_samples = sample_size / __log_entry_sz(log_offset);
709 for (i = 0; i < nr_samples; i++) {
710 s = __get_sample(samples, log_offset, i);
712 entry = (struct io_u_plat_entry *) s->val;
713 io_u_plat = entry->io_u_plat;
715 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
716 io_u_plat_before = entry_before->io_u_plat;
718 fprintf(f, "%lu, %u, %u, ", (unsigned long)s->time,
719 io_sample_ddir(s), s->bs);
720 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
721 fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat, io_u_plat_before));
723 fprintf(f, "%lu\n", (unsigned long)
724 hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
727 flist_del(&entry_before->list);
732 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
736 uint64_t i, nr_samples;
741 s = __get_sample(samples, 0, 0);
742 log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
744 nr_samples = sample_size / __log_entry_sz(log_offset);
746 for (i = 0; i < nr_samples; i++) {
747 s = __get_sample(samples, log_offset, i);
750 fprintf(f, "%lu, %lu, %u, %u\n",
751 (unsigned long) s->time,
752 (unsigned long) s->val,
753 io_sample_ddir(s), s->bs);
755 struct io_sample_offset *so = (void *) s;
757 fprintf(f, "%lu, %lu, %u, %u, %llu\n",
758 (unsigned long) s->time,
759 (unsigned long) s->val,
760 io_sample_ddir(s), s->bs,
761 (unsigned long long) so->offset);
768 struct iolog_flush_data {
769 struct workqueue_work work;
776 #define GZ_CHUNK 131072
778 static struct iolog_compress *get_new_chunk(unsigned int seq)
780 struct iolog_compress *c;
782 c = malloc(sizeof(*c));
783 INIT_FLIST_HEAD(&c->list);
784 c->buf = malloc(GZ_CHUNK);
790 static void free_chunk(struct iolog_compress *ic)
796 static int z_stream_init(z_stream *stream, int gz_hdr)
800 memset(stream, 0, sizeof(*stream));
801 stream->zalloc = Z_NULL;
802 stream->zfree = Z_NULL;
803 stream->opaque = Z_NULL;
804 stream->next_in = Z_NULL;
807 * zlib magic - add 32 for auto-detection of gz header or not,
808 * if we decide to store files in a gzip friendly format.
813 if (inflateInit2(stream, wbits) != Z_OK)
819 struct inflate_chunk_iter {
828 static void finish_chunk(z_stream *stream, FILE *f,
829 struct inflate_chunk_iter *iter)
833 ret = inflateEnd(stream);
835 log_err("fio: failed to end log inflation seq %d (%d)\n",
838 flush_samples(f, iter->buf, iter->buf_used);
841 iter->buf_size = iter->buf_used = 0;
845 * Iterative chunk inflation. Handles cases where we cross into a new
846 * sequence, doing flush finish of previous chunk if needed.
848 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
849 z_stream *stream, struct inflate_chunk_iter *iter)
853 dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
854 (unsigned long) ic->len, ic->seq);
856 if (ic->seq != iter->seq) {
858 finish_chunk(stream, f, iter);
860 z_stream_init(stream, gz_hdr);
864 stream->avail_in = ic->len;
865 stream->next_in = ic->buf;
867 if (!iter->buf_size) {
868 iter->buf_size = iter->chunk_sz;
869 iter->buf = malloc(iter->buf_size);
872 while (stream->avail_in) {
873 size_t this_out = iter->buf_size - iter->buf_used;
876 stream->avail_out = this_out;
877 stream->next_out = iter->buf + iter->buf_used;
879 err = inflate(stream, Z_NO_FLUSH);
881 log_err("fio: failed inflating log: %d\n", err);
886 iter->buf_used += this_out - stream->avail_out;
888 if (!stream->avail_out) {
889 iter->buf_size += iter->chunk_sz;
890 iter->buf = realloc(iter->buf, iter->buf_size);
894 if (err == Z_STREAM_END)
898 ret = (void *) stream->next_in - ic->buf;
900 dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
906 * Inflate stored compressed chunks, or write them directly to the log
907 * file if so instructed.
909 static int inflate_gz_chunks(struct io_log *log, FILE *f)
911 struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
914 while (!flist_empty(&log->chunk_list)) {
915 struct iolog_compress *ic;
917 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
918 flist_del(&ic->list);
920 if (log->log_gz_store) {
923 dprint(FD_COMPRESS, "log write chunk size=%lu, "
924 "seq=%u\n", (unsigned long) ic->len, ic->seq);
926 ret = fwrite(ic->buf, ic->len, 1, f);
927 if (ret != 1 || ferror(f)) {
929 log_err("fio: error writing compressed log\n");
932 inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
938 finish_chunk(&stream, f, &iter);
946 * Open compressed log file and decompress the stored chunks and
947 * write them to stdout. The chunks are stored sequentially in the
948 * file, so we iterate over them and do them one-by-one.
950 int iolog_file_inflate(const char *file)
952 struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
953 struct iolog_compress ic;
961 f = fopen(file, "r");
967 if (stat(file, &sb) < 0) {
973 ic.buf = buf = malloc(sb.st_size);
977 ret = fread(ic.buf, ic.len, 1, f);
983 } else if (ret != 1) {
984 log_err("fio: short read on reading log\n");
993 * Each chunk will return Z_STREAM_END. We don't know how many
994 * chunks are in the file, so we just keep looping and incrementing
995 * the sequence number until we have consumed the whole compressed
1002 iret = inflate_chunk(&ic, 1, stdout, &stream, &iter);
1015 finish_chunk(&stream, stdout, &iter);
1025 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1030 int iolog_file_inflate(const char *file)
1032 log_err("fio: log inflation not possible without zlib\n");
1038 void flush_log(struct io_log *log, bool do_append)
1044 f = fopen(log->filename, "w");
1046 f = fopen(log->filename, "a");
1048 perror("fopen log");
1052 buf = set_file_buffer(f);
1054 inflate_gz_chunks(log, f);
1056 while (!flist_empty(&log->io_logs)) {
1057 struct io_logs *cur_log;
1059 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1060 flist_del_init(&cur_log->list);
1062 if (log == log->td->clat_hist_log)
1063 flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1064 cur_log->nr_samples * log_entry_sz(log));
1066 flush_samples(f, cur_log->log, cur_log->nr_samples * log_entry_sz(log));
1072 clear_file_buffer(buf);
1075 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1077 if (td->flags & TD_F_COMPRESS_LOG)
1081 if (fio_trylock_file(log->filename))
1084 fio_lock_file(log->filename);
1086 if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1087 fio_send_iolog(td, log, log->filename);
1089 flush_log(log, !td->o.per_job_logs);
1091 fio_unlock_file(log->filename);
1096 size_t log_chunk_sizes(struct io_log *log)
1098 struct flist_head *entry;
1101 if (flist_empty(&log->chunk_list))
1105 pthread_mutex_lock(&log->chunk_lock);
1106 flist_for_each(entry, &log->chunk_list) {
1107 struct iolog_compress *c;
1109 c = flist_entry(entry, struct iolog_compress, list);
1112 pthread_mutex_unlock(&log->chunk_lock);
1118 static int gz_work(struct iolog_flush_data *data)
1120 struct iolog_compress *c = NULL;
1121 struct flist_head list;
1127 INIT_FLIST_HEAD(&list);
1129 memset(&stream, 0, sizeof(stream));
1130 stream.zalloc = Z_NULL;
1131 stream.zfree = Z_NULL;
1132 stream.opaque = Z_NULL;
1134 ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1136 log_err("fio: failed to init gz stream\n");
1140 seq = ++data->log->chunk_seq;
1142 stream.next_in = (void *) data->samples;
1143 stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1145 dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1146 (unsigned long) stream.avail_in, seq,
1147 data->log->filename);
1150 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1151 c = get_new_chunk(seq);
1152 stream.avail_out = GZ_CHUNK;
1153 stream.next_out = c->buf;
1154 ret = deflate(&stream, Z_NO_FLUSH);
1156 log_err("fio: deflate log (%d)\n", ret);
1161 c->len = GZ_CHUNK - stream.avail_out;
1162 flist_add_tail(&c->list, &list);
1164 } while (stream.avail_in);
1166 stream.next_out = c->buf + c->len;
1167 stream.avail_out = GZ_CHUNK - c->len;
1169 ret = deflate(&stream, Z_FINISH);
1172 * Z_BUF_ERROR is special, it just means we need more
1173 * output space. We'll handle that below. Treat any other
1176 if (ret != Z_BUF_ERROR) {
1177 log_err("fio: deflate log (%d)\n", ret);
1178 flist_del(&c->list);
1185 c->len = GZ_CHUNK - stream.avail_out;
1187 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1189 if (ret != Z_STREAM_END) {
1191 c = get_new_chunk(seq);
1192 stream.avail_out = GZ_CHUNK;
1193 stream.next_out = c->buf;
1194 ret = deflate(&stream, Z_FINISH);
1195 c->len = GZ_CHUNK - stream.avail_out;
1197 flist_add_tail(&c->list, &list);
1198 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1199 } while (ret != Z_STREAM_END);
1202 dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1204 ret = deflateEnd(&stream);
1206 log_err("fio: deflateEnd %d\n", ret);
1208 free(data->samples);
1210 if (!flist_empty(&list)) {
1211 pthread_mutex_lock(&data->log->chunk_lock);
1212 flist_splice_tail(&list, &data->log->chunk_list);
1213 pthread_mutex_unlock(&data->log->chunk_lock);
1222 while (!flist_empty(&list)) {
1223 c = flist_first_entry(list.next, struct iolog_compress, list);
1224 flist_del(&c->list);
1232 * Invoked from our compress helper thread, when logging would have exceeded
1233 * the specified memory limitation. Compresses the previously stored
1236 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1238 return gz_work(container_of(work, struct iolog_flush_data, work));
1241 static int gz_init_worker(struct submit_worker *sw)
1243 struct thread_data *td = sw->wq->td;
1245 if (!fio_option_is_set(&td->o, log_gz_cpumask))
1248 if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1249 log_err("gz: failed to set CPU affinity\n");
1256 static struct workqueue_ops log_compress_wq_ops = {
1257 .fn = gz_work_async,
1258 .init_worker_fn = gz_init_worker,
1262 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1264 if (!(td->flags & TD_F_COMPRESS_LOG))
1267 workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1271 void iolog_compress_exit(struct thread_data *td)
1273 if (!(td->flags & TD_F_COMPRESS_LOG))
1276 workqueue_exit(&td->log_compress_wq);
1280 * Queue work item to compress the existing log entries. We reset the
1281 * current log to a small size, and reference the existing log in the
1282 * data that we queue for compression. Once compression has been done,
1283 * this old log is freed. If called with finish == true, will not return
1284 * until the log compression has completed, and will flush all previous
1287 static int iolog_flush(struct io_log *log)
1289 struct iolog_flush_data *data;
1291 data = malloc(sizeof(*data));
1298 while (!flist_empty(&log->io_logs)) {
1299 struct io_logs *cur_log;
1301 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1302 flist_del_init(&cur_log->list);
1304 data->samples = cur_log->log;
1305 data->nr_samples = cur_log->nr_samples;
1316 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1318 struct iolog_flush_data *data;
1320 data = malloc(sizeof(*data));
1326 data->samples = cur_log->log;
1327 data->nr_samples = cur_log->nr_samples;
1330 cur_log->nr_samples = cur_log->max_samples = 0;
1331 cur_log->log = NULL;
1333 workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1338 static int iolog_flush(struct io_log *log)
1343 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1348 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1353 void iolog_compress_exit(struct thread_data *td)
1359 struct io_logs *iolog_cur_log(struct io_log *log)
1361 if (flist_empty(&log->io_logs))
1364 return flist_last_entry(&log->io_logs, struct io_logs, list);
1367 uint64_t iolog_nr_samples(struct io_log *iolog)
1369 struct flist_head *entry;
1372 flist_for_each(entry, &iolog->io_logs) {
1373 struct io_logs *cur_log;
1375 cur_log = flist_entry(entry, struct io_logs, list);
1376 ret += cur_log->nr_samples;
1382 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1385 return finish_log(td, log, try);
1390 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1394 if (per_unit_log(td->iops_log) != unit_log)
1397 ret = __write_log(td, td->iops_log, try);
1399 td->iops_log = NULL;
1404 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1411 ret = __write_log(td, td->slat_log, try);
1413 td->slat_log = NULL;
1418 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1425 ret = __write_log(td, td->clat_log, try);
1427 td->clat_log = NULL;
1432 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1439 ret = __write_log(td, td->clat_hist_log, try);
1441 td->clat_hist_log = NULL;
1446 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1453 ret = __write_log(td, td->lat_log, try);
1460 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1464 if (per_unit_log(td->bw_log) != unit_log)
1467 ret = __write_log(td, td->bw_log, try);
1480 CLAT_HIST_LOG_MASK = 32,
1487 int (*fn)(struct thread_data *, int, bool);
1490 static struct log_type log_types[] = {
1492 .mask = BW_LOG_MASK,
1493 .fn = write_bandw_log,
1496 .mask = LAT_LOG_MASK,
1497 .fn = write_lat_log,
1500 .mask = SLAT_LOG_MASK,
1501 .fn = write_slat_log,
1504 .mask = CLAT_LOG_MASK,
1505 .fn = write_clat_log,
1508 .mask = IOPS_LOG_MASK,
1509 .fn = write_iops_log,
1512 .mask = CLAT_HIST_LOG_MASK,
1513 .fn = write_clat_hist_log,
1517 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1519 unsigned int log_mask = 0;
1520 unsigned int log_left = ALL_LOG_NR;
1523 old_state = td_bump_runstate(td, TD_FINISHING);
1525 finalize_logs(td, unit_logs);
1528 int prev_log_left = log_left;
1530 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1531 struct log_type *lt = &log_types[i];
1534 if (!(log_mask & lt->mask)) {
1535 ret = lt->fn(td, log_left != 1, unit_logs);
1538 log_mask |= lt->mask;
1543 if (prev_log_left == log_left)
1547 td_restore_runstate(td, old_state);
1550 void fio_writeout_logs(bool unit_logs)
1552 struct thread_data *td;
1556 td_writeout_logs(td, unit_logs);