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 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->hist_msec = p->hist_msec;
588 l->hist_coarseness = p->hist_coarseness;
589 l->filename = strdup(filename);
592 if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
595 p = calloc(1, sizeof(*l->pending));
596 p->max_samples = DEF_LOG_ENTRIES;
597 p->log = calloc(p->max_samples, log_entry_sz(l));
602 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
604 INIT_FLIST_HEAD(&l->chunk_list);
606 if (l->log_gz && !p->td)
608 else if (l->log_gz || l->log_gz_store) {
609 mutex_init_pshared(&l->chunk_lock);
610 p->td->flags |= TD_F_COMPRESS_LOG;
616 #ifdef CONFIG_SETVBUF
617 static void *set_file_buffer(FILE *f)
619 size_t size = 1048576;
623 setvbuf(f, buf, _IOFBF, size);
627 static void clear_file_buffer(void *buf)
632 static void *set_file_buffer(FILE *f)
637 static void clear_file_buffer(void *buf)
642 void free_log(struct io_log *log)
644 while (!flist_empty(&log->io_logs)) {
645 struct io_logs *cur_log;
647 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
648 flist_del_init(&cur_log->list);
654 free(log->pending->log);
664 static inline unsigned long hist_sum(int j, int stride, unsigned int *io_u_plat)
669 for (k = sum = 0; k < stride; k++)
670 sum += io_u_plat[j + k];
675 void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
676 uint64_t sample_size)
680 uint64_t i, j, nr_samples;
681 unsigned int *io_u_plat;
683 int stride = 1 << hist_coarseness;
688 s = __get_sample(samples, 0, 0);
689 log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
691 nr_samples = sample_size / __log_entry_sz(log_offset);
693 for (i = 0; i < nr_samples; i++) {
694 s = __get_sample(samples, log_offset, i);
695 io_u_plat = (unsigned int *) (uintptr_t) s->val;
696 fprintf(f, "%lu, %u, %u, ", (unsigned long)s->time,
697 io_sample_ddir(s), s->bs);
698 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
699 fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat));
701 fprintf(f, "%lu\n", (unsigned long)
702 hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat));
707 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
711 uint64_t i, nr_samples;
716 s = __get_sample(samples, 0, 0);
717 log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
719 nr_samples = sample_size / __log_entry_sz(log_offset);
721 for (i = 0; i < nr_samples; i++) {
722 s = __get_sample(samples, log_offset, i);
725 fprintf(f, "%lu, %lu, %u, %u\n",
726 (unsigned long) s->time,
727 (unsigned long) s->val,
728 io_sample_ddir(s), s->bs);
730 struct io_sample_offset *so = (void *) s;
732 fprintf(f, "%lu, %lu, %u, %u, %llu\n",
733 (unsigned long) s->time,
734 (unsigned long) s->val,
735 io_sample_ddir(s), s->bs,
736 (unsigned long long) so->offset);
743 struct iolog_flush_data {
744 struct workqueue_work work;
751 #define GZ_CHUNK 131072
753 static struct iolog_compress *get_new_chunk(unsigned int seq)
755 struct iolog_compress *c;
757 c = malloc(sizeof(*c));
758 INIT_FLIST_HEAD(&c->list);
759 c->buf = malloc(GZ_CHUNK);
765 static void free_chunk(struct iolog_compress *ic)
771 static int z_stream_init(z_stream *stream, int gz_hdr)
775 memset(stream, 0, sizeof(*stream));
776 stream->zalloc = Z_NULL;
777 stream->zfree = Z_NULL;
778 stream->opaque = Z_NULL;
779 stream->next_in = Z_NULL;
782 * zlib magic - add 32 for auto-detection of gz header or not,
783 * if we decide to store files in a gzip friendly format.
788 if (inflateInit2(stream, wbits) != Z_OK)
794 struct inflate_chunk_iter {
803 static void finish_chunk(z_stream *stream, FILE *f,
804 struct inflate_chunk_iter *iter)
808 ret = inflateEnd(stream);
810 log_err("fio: failed to end log inflation seq %d (%d)\n",
813 flush_samples(f, iter->buf, iter->buf_used);
816 iter->buf_size = iter->buf_used = 0;
820 * Iterative chunk inflation. Handles cases where we cross into a new
821 * sequence, doing flush finish of previous chunk if needed.
823 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
824 z_stream *stream, struct inflate_chunk_iter *iter)
828 dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
829 (unsigned long) ic->len, ic->seq);
831 if (ic->seq != iter->seq) {
833 finish_chunk(stream, f, iter);
835 z_stream_init(stream, gz_hdr);
839 stream->avail_in = ic->len;
840 stream->next_in = ic->buf;
842 if (!iter->buf_size) {
843 iter->buf_size = iter->chunk_sz;
844 iter->buf = malloc(iter->buf_size);
847 while (stream->avail_in) {
848 size_t this_out = iter->buf_size - iter->buf_used;
851 stream->avail_out = this_out;
852 stream->next_out = iter->buf + iter->buf_used;
854 err = inflate(stream, Z_NO_FLUSH);
856 log_err("fio: failed inflating log: %d\n", err);
861 iter->buf_used += this_out - stream->avail_out;
863 if (!stream->avail_out) {
864 iter->buf_size += iter->chunk_sz;
865 iter->buf = realloc(iter->buf, iter->buf_size);
869 if (err == Z_STREAM_END)
873 ret = (void *) stream->next_in - ic->buf;
875 dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
881 * Inflate stored compressed chunks, or write them directly to the log
882 * file if so instructed.
884 static int inflate_gz_chunks(struct io_log *log, FILE *f)
886 struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
889 while (!flist_empty(&log->chunk_list)) {
890 struct iolog_compress *ic;
892 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
893 flist_del(&ic->list);
895 if (log->log_gz_store) {
898 dprint(FD_COMPRESS, "log write chunk size=%lu, "
899 "seq=%u\n", (unsigned long) ic->len, ic->seq);
901 ret = fwrite(ic->buf, ic->len, 1, f);
902 if (ret != 1 || ferror(f)) {
904 log_err("fio: error writing compressed log\n");
907 inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
913 finish_chunk(&stream, f, &iter);
921 * Open compressed log file and decompress the stored chunks and
922 * write them to stdout. The chunks are stored sequentially in the
923 * file, so we iterate over them and do them one-by-one.
925 int iolog_file_inflate(const char *file)
927 struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
928 struct iolog_compress ic;
936 f = fopen(file, "r");
942 if (stat(file, &sb) < 0) {
948 ic.buf = buf = malloc(sb.st_size);
952 ret = fread(ic.buf, ic.len, 1, f);
958 } else if (ret != 1) {
959 log_err("fio: short read on reading log\n");
968 * Each chunk will return Z_STREAM_END. We don't know how many
969 * chunks are in the file, so we just keep looping and incrementing
970 * the sequence number until we have consumed the whole compressed
977 iret = inflate_chunk(&ic, 1, stdout, &stream, &iter);
990 finish_chunk(&stream, stdout, &iter);
1000 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1005 int iolog_file_inflate(const char *file)
1007 log_err("fio: log inflation not possible without zlib\n");
1013 void flush_log(struct io_log *log, bool do_append)
1019 f = fopen(log->filename, "w");
1021 f = fopen(log->filename, "a");
1023 perror("fopen log");
1027 buf = set_file_buffer(f);
1029 inflate_gz_chunks(log, f);
1031 while (!flist_empty(&log->io_logs)) {
1032 struct io_logs *cur_log;
1034 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1035 flist_del_init(&cur_log->list);
1037 if (log == log->td->clat_hist_log)
1038 flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1039 cur_log->nr_samples * log_entry_sz(log));
1041 flush_samples(f, cur_log->log, cur_log->nr_samples * log_entry_sz(log));
1047 clear_file_buffer(buf);
1050 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1052 if (td->flags & TD_F_COMPRESS_LOG)
1056 if (fio_trylock_file(log->filename))
1059 fio_lock_file(log->filename);
1061 if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1062 fio_send_iolog(td, log, log->filename);
1064 flush_log(log, !td->o.per_job_logs);
1066 fio_unlock_file(log->filename);
1071 size_t log_chunk_sizes(struct io_log *log)
1073 struct flist_head *entry;
1076 if (flist_empty(&log->chunk_list))
1080 pthread_mutex_lock(&log->chunk_lock);
1081 flist_for_each(entry, &log->chunk_list) {
1082 struct iolog_compress *c;
1084 c = flist_entry(entry, struct iolog_compress, list);
1087 pthread_mutex_unlock(&log->chunk_lock);
1093 static int gz_work(struct iolog_flush_data *data)
1095 struct iolog_compress *c = NULL;
1096 struct flist_head list;
1102 INIT_FLIST_HEAD(&list);
1104 memset(&stream, 0, sizeof(stream));
1105 stream.zalloc = Z_NULL;
1106 stream.zfree = Z_NULL;
1107 stream.opaque = Z_NULL;
1109 ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1111 log_err("fio: failed to init gz stream\n");
1115 seq = ++data->log->chunk_seq;
1117 stream.next_in = (void *) data->samples;
1118 stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1120 dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1121 (unsigned long) stream.avail_in, seq,
1122 data->log->filename);
1125 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1126 c = get_new_chunk(seq);
1127 stream.avail_out = GZ_CHUNK;
1128 stream.next_out = c->buf;
1129 ret = deflate(&stream, Z_NO_FLUSH);
1131 log_err("fio: deflate log (%d)\n", ret);
1136 c->len = GZ_CHUNK - stream.avail_out;
1137 flist_add_tail(&c->list, &list);
1139 } while (stream.avail_in);
1141 stream.next_out = c->buf + c->len;
1142 stream.avail_out = GZ_CHUNK - c->len;
1144 ret = deflate(&stream, Z_FINISH);
1147 * Z_BUF_ERROR is special, it just means we need more
1148 * output space. We'll handle that below. Treat any other
1151 if (ret != Z_BUF_ERROR) {
1152 log_err("fio: deflate log (%d)\n", ret);
1153 flist_del(&c->list);
1160 c->len = GZ_CHUNK - stream.avail_out;
1162 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1164 if (ret != Z_STREAM_END) {
1166 c = get_new_chunk(seq);
1167 stream.avail_out = GZ_CHUNK;
1168 stream.next_out = c->buf;
1169 ret = deflate(&stream, Z_FINISH);
1170 c->len = GZ_CHUNK - stream.avail_out;
1172 flist_add_tail(&c->list, &list);
1173 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1174 } while (ret != Z_STREAM_END);
1177 dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1179 ret = deflateEnd(&stream);
1181 log_err("fio: deflateEnd %d\n", ret);
1183 free(data->samples);
1185 if (!flist_empty(&list)) {
1186 pthread_mutex_lock(&data->log->chunk_lock);
1187 flist_splice_tail(&list, &data->log->chunk_list);
1188 pthread_mutex_unlock(&data->log->chunk_lock);
1197 while (!flist_empty(&list)) {
1198 c = flist_first_entry(list.next, struct iolog_compress, list);
1199 flist_del(&c->list);
1207 * Invoked from our compress helper thread, when logging would have exceeded
1208 * the specified memory limitation. Compresses the previously stored
1211 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1213 return gz_work(container_of(work, struct iolog_flush_data, work));
1216 static int gz_init_worker(struct submit_worker *sw)
1218 struct thread_data *td = sw->wq->td;
1220 if (!fio_option_is_set(&td->o, log_gz_cpumask))
1223 if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1224 log_err("gz: failed to set CPU affinity\n");
1231 static struct workqueue_ops log_compress_wq_ops = {
1232 .fn = gz_work_async,
1233 .init_worker_fn = gz_init_worker,
1237 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1239 if (!(td->flags & TD_F_COMPRESS_LOG))
1242 workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1246 void iolog_compress_exit(struct thread_data *td)
1248 if (!(td->flags & TD_F_COMPRESS_LOG))
1251 workqueue_exit(&td->log_compress_wq);
1255 * Queue work item to compress the existing log entries. We reset the
1256 * current log to a small size, and reference the existing log in the
1257 * data that we queue for compression. Once compression has been done,
1258 * this old log is freed. If called with finish == true, will not return
1259 * until the log compression has completed, and will flush all previous
1262 static int iolog_flush(struct io_log *log)
1264 struct iolog_flush_data *data;
1266 data = malloc(sizeof(*data));
1273 while (!flist_empty(&log->io_logs)) {
1274 struct io_logs *cur_log;
1276 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1277 flist_del_init(&cur_log->list);
1279 data->samples = cur_log->log;
1280 data->nr_samples = cur_log->nr_samples;
1291 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1293 struct iolog_flush_data *data;
1295 data = malloc(sizeof(*data));
1301 data->samples = cur_log->log;
1302 data->nr_samples = cur_log->nr_samples;
1305 cur_log->nr_samples = cur_log->max_samples = 0;
1306 cur_log->log = NULL;
1308 workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1313 static int iolog_flush(struct io_log *log)
1318 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1323 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1328 void iolog_compress_exit(struct thread_data *td)
1334 struct io_logs *iolog_cur_log(struct io_log *log)
1336 if (flist_empty(&log->io_logs))
1339 return flist_last_entry(&log->io_logs, struct io_logs, list);
1342 uint64_t iolog_nr_samples(struct io_log *iolog)
1344 struct flist_head *entry;
1347 flist_for_each(entry, &iolog->io_logs) {
1348 struct io_logs *cur_log;
1350 cur_log = flist_entry(entry, struct io_logs, list);
1351 ret += cur_log->nr_samples;
1357 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1360 return finish_log(td, log, try);
1365 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1369 if (per_unit_log(td->iops_log) != unit_log)
1372 ret = __write_log(td, td->iops_log, try);
1374 td->iops_log = NULL;
1379 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1386 ret = __write_log(td, td->slat_log, try);
1388 td->slat_log = NULL;
1393 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1400 ret = __write_log(td, td->clat_log, try);
1402 td->clat_log = NULL;
1407 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1414 ret = __write_log(td, td->clat_hist_log, try);
1416 td->clat_hist_log = NULL;
1421 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1428 ret = __write_log(td, td->lat_log, try);
1435 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1439 if (per_unit_log(td->bw_log) != unit_log)
1442 ret = __write_log(td, td->bw_log, try);
1455 CLAT_HIST_LOG_MASK = 32,
1462 int (*fn)(struct thread_data *, int, bool);
1465 static struct log_type log_types[] = {
1467 .mask = BW_LOG_MASK,
1468 .fn = write_bandw_log,
1471 .mask = LAT_LOG_MASK,
1472 .fn = write_lat_log,
1475 .mask = SLAT_LOG_MASK,
1476 .fn = write_slat_log,
1479 .mask = CLAT_LOG_MASK,
1480 .fn = write_clat_log,
1483 .mask = IOPS_LOG_MASK,
1484 .fn = write_iops_log,
1487 .mask = CLAT_HIST_LOG_MASK,
1488 .fn = write_clat_hist_log,
1492 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1494 unsigned int log_mask = 0;
1495 unsigned int log_left = ALL_LOG_NR;
1498 old_state = td_bump_runstate(td, TD_FINISHING);
1500 finalize_logs(td, unit_logs);
1503 int prev_log_left = log_left;
1505 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1506 struct log_type *lt = &log_types[i];
1509 if (!(log_mask & lt->mask)) {
1510 ret = lt->fn(td, log_left != 1, unit_logs);
1513 log_mask |= lt->mask;
1518 if (prev_log_left == log_left)
1522 td_restore_runstate(td, old_state);
1525 void fio_writeout_logs(bool unit_logs)
1527 struct thread_data *td;
1531 td_writeout_logs(td, unit_logs);