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 int hist_sum(int j, int stride, unsigned int *io_u_plat) {
666 for (int k = 0; k < stride; k++) {
667 sum += io_u_plat[j + k];
672 void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
673 uint64_t sample_size)
677 uint64_t i, j, nr_samples;
678 unsigned int *io_u_plat;
680 int stride = 1 << hist_coarseness;
685 s = __get_sample(samples, 0, 0);
686 log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
688 nr_samples = sample_size / __log_entry_sz(log_offset);
690 for (i = 0; i < nr_samples; i++) {
691 s = __get_sample(samples, log_offset, i);
692 io_u_plat = (unsigned int *)(s->val);
693 fprintf(f, "%lu, %u, %u, ", (unsigned long)s->time,
694 io_sample_ddir(s), s->bs);
695 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
696 fprintf(f, "%lu, ", (unsigned long) hist_sum(j, stride, io_u_plat));
698 fprintf(f, "%lu\n", (unsigned long)
699 hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat));
704 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
708 uint64_t i, nr_samples;
713 s = __get_sample(samples, 0, 0);
714 log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
716 nr_samples = sample_size / __log_entry_sz(log_offset);
718 for (i = 0; i < nr_samples; i++) {
719 s = __get_sample(samples, log_offset, i);
722 fprintf(f, "%lu, %lu, %u, %u\n",
723 (unsigned long) s->time,
724 (unsigned long) s->val,
725 io_sample_ddir(s), s->bs);
727 struct io_sample_offset *so = (void *) s;
729 fprintf(f, "%lu, %lu, %u, %u, %llu\n",
730 (unsigned long) s->time,
731 (unsigned long) s->val,
732 io_sample_ddir(s), s->bs,
733 (unsigned long long) so->offset);
740 struct iolog_flush_data {
741 struct workqueue_work work;
748 #define GZ_CHUNK 131072
750 static struct iolog_compress *get_new_chunk(unsigned int seq)
752 struct iolog_compress *c;
754 c = malloc(sizeof(*c));
755 INIT_FLIST_HEAD(&c->list);
756 c->buf = malloc(GZ_CHUNK);
762 static void free_chunk(struct iolog_compress *ic)
768 static int z_stream_init(z_stream *stream, int gz_hdr)
772 memset(stream, 0, sizeof(*stream));
773 stream->zalloc = Z_NULL;
774 stream->zfree = Z_NULL;
775 stream->opaque = Z_NULL;
776 stream->next_in = Z_NULL;
779 * zlib magic - add 32 for auto-detection of gz header or not,
780 * if we decide to store files in a gzip friendly format.
785 if (inflateInit2(stream, wbits) != Z_OK)
791 struct inflate_chunk_iter {
800 static void finish_chunk(z_stream *stream, FILE *f,
801 struct inflate_chunk_iter *iter)
805 ret = inflateEnd(stream);
807 log_err("fio: failed to end log inflation seq %d (%d)\n",
810 flush_samples(f, iter->buf, iter->buf_used);
813 iter->buf_size = iter->buf_used = 0;
817 * Iterative chunk inflation. Handles cases where we cross into a new
818 * sequence, doing flush finish of previous chunk if needed.
820 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
821 z_stream *stream, struct inflate_chunk_iter *iter)
825 dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
826 (unsigned long) ic->len, ic->seq);
828 if (ic->seq != iter->seq) {
830 finish_chunk(stream, f, iter);
832 z_stream_init(stream, gz_hdr);
836 stream->avail_in = ic->len;
837 stream->next_in = ic->buf;
839 if (!iter->buf_size) {
840 iter->buf_size = iter->chunk_sz;
841 iter->buf = malloc(iter->buf_size);
844 while (stream->avail_in) {
845 size_t this_out = iter->buf_size - iter->buf_used;
848 stream->avail_out = this_out;
849 stream->next_out = iter->buf + iter->buf_used;
851 err = inflate(stream, Z_NO_FLUSH);
853 log_err("fio: failed inflating log: %d\n", err);
858 iter->buf_used += this_out - stream->avail_out;
860 if (!stream->avail_out) {
861 iter->buf_size += iter->chunk_sz;
862 iter->buf = realloc(iter->buf, iter->buf_size);
866 if (err == Z_STREAM_END)
870 ret = (void *) stream->next_in - ic->buf;
872 dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
878 * Inflate stored compressed chunks, or write them directly to the log
879 * file if so instructed.
881 static int inflate_gz_chunks(struct io_log *log, FILE *f)
883 struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
886 while (!flist_empty(&log->chunk_list)) {
887 struct iolog_compress *ic;
889 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
890 flist_del(&ic->list);
892 if (log->log_gz_store) {
895 dprint(FD_COMPRESS, "log write chunk size=%lu, "
896 "seq=%u\n", (unsigned long) ic->len, ic->seq);
898 ret = fwrite(ic->buf, ic->len, 1, f);
899 if (ret != 1 || ferror(f)) {
901 log_err("fio: error writing compressed log\n");
904 inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
910 finish_chunk(&stream, f, &iter);
918 * Open compressed log file and decompress the stored chunks and
919 * write them to stdout. The chunks are stored sequentially in the
920 * file, so we iterate over them and do them one-by-one.
922 int iolog_file_inflate(const char *file)
924 struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
925 struct iolog_compress ic;
933 f = fopen(file, "r");
939 if (stat(file, &sb) < 0) {
945 ic.buf = buf = malloc(sb.st_size);
949 ret = fread(ic.buf, ic.len, 1, f);
955 } else if (ret != 1) {
956 log_err("fio: short read on reading log\n");
965 * Each chunk will return Z_STREAM_END. We don't know how many
966 * chunks are in the file, so we just keep looping and incrementing
967 * the sequence number until we have consumed the whole compressed
974 iret = inflate_chunk(&ic, 1, stdout, &stream, &iter);
987 finish_chunk(&stream, stdout, &iter);
997 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1002 int iolog_file_inflate(const char *file)
1004 log_err("fio: log inflation not possible without zlib\n");
1010 void flush_log(struct io_log *log, bool do_append)
1016 f = fopen(log->filename, "w");
1018 f = fopen(log->filename, "a");
1020 perror("fopen log");
1024 buf = set_file_buffer(f);
1026 inflate_gz_chunks(log, f);
1028 while (!flist_empty(&log->io_logs)) {
1029 struct io_logs *cur_log;
1031 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1032 flist_del_init(&cur_log->list);
1034 if (log == log->td->clat_hist_log)
1035 flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1036 cur_log->nr_samples * log_entry_sz(log));
1038 flush_samples(f, cur_log->log, cur_log->nr_samples * log_entry_sz(log));
1044 clear_file_buffer(buf);
1047 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1049 if (td->flags & TD_F_COMPRESS_LOG)
1053 if (fio_trylock_file(log->filename))
1056 fio_lock_file(log->filename);
1058 if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1059 fio_send_iolog(td, log, log->filename);
1061 flush_log(log, !td->o.per_job_logs);
1063 fio_unlock_file(log->filename);
1068 size_t log_chunk_sizes(struct io_log *log)
1070 struct flist_head *entry;
1073 if (flist_empty(&log->chunk_list))
1077 pthread_mutex_lock(&log->chunk_lock);
1078 flist_for_each(entry, &log->chunk_list) {
1079 struct iolog_compress *c;
1081 c = flist_entry(entry, struct iolog_compress, list);
1084 pthread_mutex_unlock(&log->chunk_lock);
1090 static int gz_work(struct iolog_flush_data *data)
1092 struct iolog_compress *c = NULL;
1093 struct flist_head list;
1099 INIT_FLIST_HEAD(&list);
1101 memset(&stream, 0, sizeof(stream));
1102 stream.zalloc = Z_NULL;
1103 stream.zfree = Z_NULL;
1104 stream.opaque = Z_NULL;
1106 ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1108 log_err("fio: failed to init gz stream\n");
1112 seq = ++data->log->chunk_seq;
1114 stream.next_in = (void *) data->samples;
1115 stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1117 dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1118 (unsigned long) stream.avail_in, seq,
1119 data->log->filename);
1122 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1123 c = get_new_chunk(seq);
1124 stream.avail_out = GZ_CHUNK;
1125 stream.next_out = c->buf;
1126 ret = deflate(&stream, Z_NO_FLUSH);
1128 log_err("fio: deflate log (%d)\n", ret);
1133 c->len = GZ_CHUNK - stream.avail_out;
1134 flist_add_tail(&c->list, &list);
1136 } while (stream.avail_in);
1138 stream.next_out = c->buf + c->len;
1139 stream.avail_out = GZ_CHUNK - c->len;
1141 ret = deflate(&stream, Z_FINISH);
1144 * Z_BUF_ERROR is special, it just means we need more
1145 * output space. We'll handle that below. Treat any other
1148 if (ret != Z_BUF_ERROR) {
1149 log_err("fio: deflate log (%d)\n", ret);
1150 flist_del(&c->list);
1157 c->len = GZ_CHUNK - stream.avail_out;
1159 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1161 if (ret != Z_STREAM_END) {
1163 c = get_new_chunk(seq);
1164 stream.avail_out = GZ_CHUNK;
1165 stream.next_out = c->buf;
1166 ret = deflate(&stream, Z_FINISH);
1167 c->len = GZ_CHUNK - stream.avail_out;
1169 flist_add_tail(&c->list, &list);
1170 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
1171 } while (ret != Z_STREAM_END);
1174 dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1176 ret = deflateEnd(&stream);
1178 log_err("fio: deflateEnd %d\n", ret);
1180 free(data->samples);
1182 if (!flist_empty(&list)) {
1183 pthread_mutex_lock(&data->log->chunk_lock);
1184 flist_splice_tail(&list, &data->log->chunk_list);
1185 pthread_mutex_unlock(&data->log->chunk_lock);
1194 while (!flist_empty(&list)) {
1195 c = flist_first_entry(list.next, struct iolog_compress, list);
1196 flist_del(&c->list);
1204 * Invoked from our compress helper thread, when logging would have exceeded
1205 * the specified memory limitation. Compresses the previously stored
1208 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1210 return gz_work(container_of(work, struct iolog_flush_data, work));
1213 static int gz_init_worker(struct submit_worker *sw)
1215 struct thread_data *td = sw->wq->td;
1217 if (!fio_option_is_set(&td->o, log_gz_cpumask))
1220 if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1221 log_err("gz: failed to set CPU affinity\n");
1228 static struct workqueue_ops log_compress_wq_ops = {
1229 .fn = gz_work_async,
1230 .init_worker_fn = gz_init_worker,
1234 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1236 if (!(td->flags & TD_F_COMPRESS_LOG))
1239 workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1243 void iolog_compress_exit(struct thread_data *td)
1245 if (!(td->flags & TD_F_COMPRESS_LOG))
1248 workqueue_exit(&td->log_compress_wq);
1252 * Queue work item to compress the existing log entries. We reset the
1253 * current log to a small size, and reference the existing log in the
1254 * data that we queue for compression. Once compression has been done,
1255 * this old log is freed. If called with finish == true, will not return
1256 * until the log compression has completed, and will flush all previous
1259 static int iolog_flush(struct io_log *log)
1261 struct iolog_flush_data *data;
1263 data = malloc(sizeof(*data));
1270 while (!flist_empty(&log->io_logs)) {
1271 struct io_logs *cur_log;
1273 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1274 flist_del_init(&cur_log->list);
1276 data->samples = cur_log->log;
1277 data->nr_samples = cur_log->nr_samples;
1288 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1290 struct iolog_flush_data *data;
1292 data = malloc(sizeof(*data));
1298 data->samples = cur_log->log;
1299 data->nr_samples = cur_log->nr_samples;
1302 cur_log->nr_samples = cur_log->max_samples = 0;
1303 cur_log->log = NULL;
1305 workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1310 static int iolog_flush(struct io_log *log)
1315 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1320 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1325 void iolog_compress_exit(struct thread_data *td)
1331 struct io_logs *iolog_cur_log(struct io_log *log)
1333 if (flist_empty(&log->io_logs))
1336 return flist_last_entry(&log->io_logs, struct io_logs, list);
1339 uint64_t iolog_nr_samples(struct io_log *iolog)
1341 struct flist_head *entry;
1344 flist_for_each(entry, &iolog->io_logs) {
1345 struct io_logs *cur_log;
1347 cur_log = flist_entry(entry, struct io_logs, list);
1348 ret += cur_log->nr_samples;
1354 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1357 return finish_log(td, log, try);
1362 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1366 if (per_unit_log(td->iops_log) != unit_log)
1369 ret = __write_log(td, td->iops_log, try);
1371 td->iops_log = NULL;
1376 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1383 ret = __write_log(td, td->slat_log, try);
1385 td->slat_log = NULL;
1390 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1397 ret = __write_log(td, td->clat_log, try);
1399 td->clat_log = NULL;
1404 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1411 ret = __write_log(td, td->clat_hist_log, try);
1413 td->clat_hist_log = NULL;
1418 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1425 ret = __write_log(td, td->lat_log, try);
1432 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1436 if (per_unit_log(td->bw_log) != unit_log)
1439 ret = __write_log(td, td->bw_log, try);
1452 CLAT_HIST_LOG_MASK = 32,
1459 int (*fn)(struct thread_data *, int, bool);
1462 static struct log_type log_types[] = {
1464 .mask = BW_LOG_MASK,
1465 .fn = write_bandw_log,
1468 .mask = LAT_LOG_MASK,
1469 .fn = write_lat_log,
1472 .mask = SLAT_LOG_MASK,
1473 .fn = write_slat_log,
1476 .mask = CLAT_LOG_MASK,
1477 .fn = write_clat_log,
1480 .mask = IOPS_LOG_MASK,
1481 .fn = write_iops_log,
1484 .mask = CLAT_HIST_LOG_MASK,
1485 .fn = write_clat_hist_log,
1489 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1491 unsigned int log_mask = 0;
1492 unsigned int log_left = ALL_LOG_NR;
1495 old_state = td_bump_runstate(td, TD_FINISHING);
1497 finalize_logs(td, unit_logs);
1500 int prev_log_left = log_left;
1502 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1503 struct log_type *lt = &log_types[i];
1506 if (!(log_mask & lt->mask)) {
1507 ret = lt->fn(td, log_left != 1, unit_logs);
1510 log_mask |= lt->mask;
1515 if (prev_log_left == log_left)
1519 td_restore_runstate(td, old_state);
1522 void fio_writeout_logs(bool unit_logs)
1524 struct thread_data *td;
1528 td_writeout_logs(td, unit_logs);