2 * Code related to writing an iolog of what a thread is doing, and to
3 * later read that back and replay
22 #include "lib/roundup.h"
24 #include <netinet/in.h>
25 #include <netinet/tcp.h>
26 #include <arpa/inet.h>
28 #include <sys/socket.h>
31 static int iolog_flush(struct io_log *log);
33 static const char iolog_ver2[] = "fio version 2 iolog";
35 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
37 flist_add_tail(&ipo->list, &td->io_log_list);
38 td->total_io_size += ipo->len;
41 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
43 if (!td->o.write_iolog_file)
46 fprintf(td->iolog_f, "%s %s %llu %llu\n", io_u->file->file_name,
47 io_ddir_name(io_u->ddir),
48 io_u->offset, io_u->buflen);
51 void log_file(struct thread_data *td, struct fio_file *f,
52 enum file_log_act what)
54 const char *act[] = { "add", "open", "close" };
58 if (!td->o.write_iolog_file)
63 * this happens on the pre-open/close done before the job starts
68 fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
71 static void iolog_delay(struct thread_data *td, unsigned long delay)
73 uint64_t usec = utime_since_now(&td->last_issue);
74 unsigned long orig_delay = delay;
78 if (delay < td->time_offset) {
83 delay -= td->time_offset;
89 fio_gettime(&ts, NULL);
90 while (delay && !td->terminate) {
92 if (this_delay > 500000)
95 usec_sleep(td, this_delay);
99 usec = utime_since_now(&ts);
100 if (usec > orig_delay)
101 td->time_offset = usec - orig_delay;
106 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
114 if (ipo->ddir != DDIR_INVAL)
117 f = td->files[ipo->fileno];
119 switch (ipo->file_action) {
120 case FIO_LOG_OPEN_FILE:
121 if (td->o.replay_redirect && fio_file_open(f)) {
122 dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
126 ret = td_io_open_file(td, f);
129 td_verror(td, ret, "iolog open file");
131 case FIO_LOG_CLOSE_FILE:
132 td_io_close_file(td, f);
134 case FIO_LOG_UNLINK_FILE:
135 td_io_unlink_file(td, f);
138 log_err("fio: bad file action %d\n", ipo->file_action);
145 static bool read_iolog2(struct thread_data *td);
147 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
149 struct io_piece *ipo;
150 unsigned long elapsed;
152 while (!flist_empty(&td->io_log_list)) {
155 if (!td->io_log_blktrace && td->o.read_iolog_chunked) {
156 if (td->io_log_checkmark == td->io_log_current) {
157 if (!read_iolog2(td))
160 td->io_log_current--;
162 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
163 flist_del(&ipo->list);
164 remove_trim_entry(td, ipo);
166 ret = ipo_special(td, ipo);
170 } else if (ret > 0) {
175 io_u->ddir = ipo->ddir;
176 if (ipo->ddir != DDIR_WAIT) {
177 io_u->offset = ipo->offset;
178 io_u->verify_offset = ipo->offset;
179 io_u->buflen = ipo->len;
180 io_u->file = td->files[ipo->fileno];
181 get_file(io_u->file);
182 dprint(FD_IO, "iolog: get %llu/%llu/%s\n", io_u->offset,
183 io_u->buflen, io_u->file->file_name);
185 iolog_delay(td, ipo->delay);
187 elapsed = mtime_since_genesis();
188 if (ipo->delay > elapsed)
189 usec_sleep(td, (ipo->delay - elapsed) * 1000);
194 if (io_u->ddir != DDIR_WAIT)
202 void prune_io_piece_log(struct thread_data *td)
204 struct io_piece *ipo;
205 struct fio_rb_node *n;
207 while ((n = rb_first(&td->io_hist_tree)) != NULL) {
208 ipo = rb_entry(n, struct io_piece, rb_node);
209 rb_erase(n, &td->io_hist_tree);
210 remove_trim_entry(td, ipo);
215 while (!flist_empty(&td->io_hist_list)) {
216 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
217 flist_del(&ipo->list);
218 remove_trim_entry(td, ipo);
225 * log a successful write, so we can unwind the log for verify
227 void log_io_piece(struct thread_data *td, struct io_u *io_u)
229 struct fio_rb_node **p, *parent;
230 struct io_piece *ipo, *__ipo;
232 ipo = calloc(1, sizeof(struct io_piece));
234 ipo->file = io_u->file;
235 ipo->offset = io_u->offset;
236 ipo->len = io_u->buflen;
237 ipo->numberio = io_u->numberio;
238 ipo->flags = IP_F_IN_FLIGHT;
242 if (io_u_should_trim(td, io_u)) {
243 flist_add_tail(&ipo->trim_list, &td->trim_list);
248 * Only sort writes if we don't have a random map in which case we need
249 * to check for duplicate blocks and drop the old one, which we rely on
250 * the rb insert/lookup for handling.
252 if (file_randommap(td, ipo->file)) {
253 INIT_FLIST_HEAD(&ipo->list);
254 flist_add_tail(&ipo->list, &td->io_hist_list);
255 ipo->flags |= IP_F_ONLIST;
260 RB_CLEAR_NODE(&ipo->rb_node);
263 * Sort the entry into the verification list
266 p = &td->io_hist_tree.rb_node;
272 __ipo = rb_entry(parent, struct io_piece, rb_node);
273 if (ipo->file < __ipo->file)
275 else if (ipo->file > __ipo->file)
277 else if (ipo->offset < __ipo->offset) {
279 overlap = ipo->offset + ipo->len > __ipo->offset;
281 else if (ipo->offset > __ipo->offset) {
283 overlap = __ipo->offset + __ipo->len > ipo->offset;
289 dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
290 __ipo->offset, __ipo->len,
291 ipo->offset, ipo->len);
293 rb_erase(parent, &td->io_hist_tree);
294 remove_trim_entry(td, __ipo);
295 if (!(__ipo->flags & IP_F_IN_FLIGHT))
301 rb_link_node(&ipo->rb_node, parent, p);
302 rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
303 ipo->flags |= IP_F_ONRB;
307 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
309 struct io_piece *ipo = io_u->ipo;
311 if (td->ts.nr_block_infos) {
312 uint32_t *info = io_u_block_info(td, io_u);
313 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
314 if (io_u->ddir == DDIR_TRIM)
315 *info = BLOCK_INFO_SET_STATE(*info,
316 BLOCK_STATE_TRIM_FAILURE);
317 else if (io_u->ddir == DDIR_WRITE)
318 *info = BLOCK_INFO_SET_STATE(*info,
319 BLOCK_STATE_WRITE_FAILURE);
326 if (ipo->flags & IP_F_ONRB)
327 rb_erase(&ipo->rb_node, &td->io_hist_tree);
328 else if (ipo->flags & IP_F_ONLIST)
329 flist_del(&ipo->list);
336 void trim_io_piece(const struct io_u *io_u)
338 struct io_piece *ipo = io_u->ipo;
343 ipo->len = io_u->xfer_buflen - io_u->resid;
346 void write_iolog_close(struct thread_data *td)
355 td->iolog_buf = NULL;
358 static int64_t iolog_items_to_fetch(struct thread_data *td)
363 int64_t items_to_fetch;
365 if (!td->io_log_highmark)
369 fio_gettime(&now, NULL);
370 elapsed = ntime_since(&td->io_log_highmark_time, &now);
372 for_1s = (td->io_log_highmark - td->io_log_current) * 1000000000 / elapsed;
373 items_to_fetch = for_1s - td->io_log_current;
374 if (items_to_fetch < 0)
379 td->io_log_highmark = td->io_log_current + items_to_fetch;
380 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
381 fio_gettime(&td->io_log_highmark_time, NULL);
383 return items_to_fetch;
387 * Read version 2 iolog data. It is enhanced to include per-file logging,
390 static bool read_iolog2(struct thread_data *td)
392 unsigned long long offset;
394 int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
395 char *rfname, *fname, *act;
398 bool realloc = false;
399 int64_t items_to_fetch = 0;
401 if (td->o.read_iolog_chunked) {
402 items_to_fetch = iolog_items_to_fetch(td);
408 * Read in the read iolog and store it, reuse the infrastructure
409 * for doing verifications.
412 rfname = fname = malloc(256+16);
413 act = malloc(256+16);
415 reads = writes = waits = 0;
416 while ((p = fgets(str, 4096, td->io_log_rfile)) != NULL) {
417 struct io_piece *ipo;
420 r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
423 if (td->o.replay_redirect)
424 fname = td->o.replay_redirect;
430 if (!strcmp(act, "wait"))
432 else if (!strcmp(act, "read"))
434 else if (!strcmp(act, "write"))
436 else if (!strcmp(act, "sync"))
438 else if (!strcmp(act, "datasync"))
440 else if (!strcmp(act, "trim"))
443 log_err("fio: bad iolog file action: %s\n",
447 fileno = get_fileno(td, fname);
450 if (!strcmp(act, "add")) {
451 if (td->o.replay_redirect &&
452 get_fileno(td, fname) != -1) {
453 dprint(FD_FILE, "iolog: ignoring"
454 " re-add of file %s\n", fname);
456 fileno = add_file(td, fname, td->subjob_number, 1);
457 file_action = FIO_LOG_ADD_FILE;
460 } else if (!strcmp(act, "open")) {
461 fileno = get_fileno(td, fname);
462 file_action = FIO_LOG_OPEN_FILE;
463 } else if (!strcmp(act, "close")) {
464 fileno = get_fileno(td, fname);
465 file_action = FIO_LOG_CLOSE_FILE;
467 log_err("fio: bad iolog file action: %s\n",
472 log_err("bad iolog2: %s\n", p);
478 else if (rw == DDIR_WRITE) {
480 * Don't add a write for ro mode
485 } else if (rw == DDIR_WAIT) {
489 } else if (rw == DDIR_INVAL) {
490 } else if (!ddir_sync(rw)) {
491 log_err("bad ddir: %d\n", rw);
498 ipo = calloc(1, sizeof(*ipo));
501 if (rw == DDIR_WAIT) {
504 if (td->o.replay_scale)
505 ipo->offset = offset / td->o.replay_scale;
507 ipo->offset = offset;
508 ipo_bytes_align(td->o.replay_align, ipo);
511 if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw]) {
513 td->o.max_bs[rw] = bytes;
515 ipo->fileno = fileno;
516 ipo->file_action = file_action;
520 queue_io_piece(td, ipo);
522 if (td->o.read_iolog_chunked) {
523 td->io_log_current++;
525 if (items_to_fetch == 0)
534 if (td->o.read_iolog_chunked) {
535 td->io_log_highmark = td->io_log_current;
536 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
537 fio_gettime(&td->io_log_highmark_time, NULL);
540 if (writes && read_only) {
541 log_err("fio: <%s> skips replay of %d writes due to"
542 " read-only\n", td->o.name, writes);
546 if (td->o.read_iolog_chunked) {
547 if (td->io_log_current == 0) {
550 td->o.td_ddir = TD_DDIR_RW;
551 if (realloc && td->orig_buffer)
555 init_io_u_buffers(td);
560 if (!reads && !writes && !waits)
562 else if (reads && !writes)
563 td->o.td_ddir = TD_DDIR_READ;
564 else if (!reads && writes)
565 td->o.td_ddir = TD_DDIR_WRITE;
567 td->o.td_ddir = TD_DDIR_RW;
572 static bool is_socket(const char *path)
577 r = stat(path, &buf);
581 return S_ISSOCK(buf.st_mode);
584 static int open_socket(const char *path)
586 struct sockaddr_un addr;
589 fd = socket(AF_UNIX, SOCK_STREAM, 0);
593 addr.sun_family = AF_UNIX;
594 if (snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", path) >=
595 sizeof(addr.sun_path)) {
596 log_err("%s: path name %s is too long for a Unix socket\n",
600 ret = connect(fd, (const struct sockaddr *)&addr, strlen(path) + sizeof(addr.sun_family));
609 * open iolog, check version, and call appropriate parser
611 static bool init_iolog_read(struct thread_data *td, char *fname)
613 char buffer[256], *p;
616 dprint(FD_IO, "iolog: name=%s\n", fname);
618 if (is_socket(fname)) {
621 fd = open_socket(fname);
624 } else if (!strcmp(fname, "-")) {
627 f = fopen(fname, "r");
632 perror("fopen read iolog");
636 p = fgets(buffer, sizeof(buffer), f);
638 td_verror(td, errno, "iolog read");
639 log_err("fio: unable to read iolog\n");
645 * version 2 of the iolog stores a specific string as the
646 * first line, check for that
648 if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2))) {
649 free_release_files(td);
650 td->io_log_rfile = f;
651 return read_iolog2(td);
654 log_err("fio: iolog version 1 is no longer supported\n");
660 * Set up a log for storing io patterns.
662 static bool init_iolog_write(struct thread_data *td)
668 f = fopen(td->o.write_iolog_file, "a");
670 perror("fopen write iolog");
675 * That's it for writing, setup a log buffer and we're done.
678 td->iolog_buf = malloc(8192);
679 setvbuf(f, td->iolog_buf, _IOFBF, 8192);
682 * write our version line
684 if (fprintf(f, "%s\n", iolog_ver2) < 0) {
685 perror("iolog init\n");
690 * add all known files
692 for_each_file(td, ff, i)
693 log_file(td, ff, FIO_LOG_ADD_FILE);
698 bool init_iolog(struct thread_data *td)
702 if (td->o.read_iolog_file) {
704 char * fname = get_name_by_idx(td->o.read_iolog_file, td->subjob_number);
707 * Check if it's a blktrace file and load that if possible.
708 * Otherwise assume it's a normal log file and load that.
710 if (is_blktrace(fname, &need_swap)) {
711 td->io_log_blktrace = 1;
712 ret = load_blktrace(td, fname, need_swap);
714 td->io_log_blktrace = 0;
715 ret = init_iolog_read(td, fname);
717 } else if (td->o.write_iolog_file)
718 ret = init_iolog_write(td);
723 td_verror(td, EINVAL, "failed initializing iolog");
728 void setup_log(struct io_log **log, struct log_params *p,
729 const char *filename)
733 struct io_u_plat_entry *entry;
734 struct flist_head *list;
736 l = scalloc(1, sizeof(*l));
737 INIT_FLIST_HEAD(&l->io_logs);
738 l->log_type = p->log_type;
739 l->log_offset = p->log_offset;
740 l->log_gz = p->log_gz;
741 l->log_gz_store = p->log_gz_store;
742 l->avg_msec = p->avg_msec;
743 l->hist_msec = p->hist_msec;
744 l->hist_coarseness = p->hist_coarseness;
745 l->filename = strdup(filename);
748 /* Initialize histogram lists for each r/w direction,
749 * with initial io_u_plat of all zeros:
751 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
752 list = &l->hist_window[i].list;
753 INIT_FLIST_HEAD(list);
754 entry = calloc(1, sizeof(struct io_u_plat_entry));
755 flist_add(&entry->list, list);
758 if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
759 unsigned int def_samples = DEF_LOG_ENTRIES;
762 __p = calloc(1, sizeof(*l->pending));
763 if (l->td->o.iodepth > DEF_LOG_ENTRIES)
764 def_samples = roundup_pow2(l->td->o.iodepth);
765 __p->max_samples = def_samples;
766 __p->log = calloc(__p->max_samples, log_entry_sz(l));
771 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
773 INIT_FLIST_HEAD(&l->chunk_list);
775 if (l->log_gz && !p->td)
777 else if (l->log_gz || l->log_gz_store) {
778 mutex_init_pshared(&l->chunk_lock);
779 mutex_init_pshared(&l->deferred_free_lock);
780 p->td->flags |= TD_F_COMPRESS_LOG;
786 #ifdef CONFIG_SETVBUF
787 static void *set_file_buffer(FILE *f)
789 size_t size = 1048576;
793 setvbuf(f, buf, _IOFBF, size);
797 static void clear_file_buffer(void *buf)
802 static void *set_file_buffer(FILE *f)
807 static void clear_file_buffer(void *buf)
812 void free_log(struct io_log *log)
814 while (!flist_empty(&log->io_logs)) {
815 struct io_logs *cur_log;
817 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
818 flist_del_init(&cur_log->list);
824 free(log->pending->log);
834 uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
835 uint64_t *io_u_plat_last)
840 if (io_u_plat_last) {
841 for (k = sum = 0; k < stride; k++)
842 sum += io_u_plat[j + k] - io_u_plat_last[j + k];
844 for (k = sum = 0; k < stride; k++)
845 sum += io_u_plat[j + k];
851 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
852 uint64_t sample_size)
856 uint64_t i, j, nr_samples;
857 struct io_u_plat_entry *entry, *entry_before;
859 uint64_t *io_u_plat_before;
861 int stride = 1 << hist_coarseness;
866 s = __get_sample(samples, 0, 0);
867 log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
869 nr_samples = sample_size / __log_entry_sz(log_offset);
871 for (i = 0; i < nr_samples; i++) {
872 s = __get_sample(samples, log_offset, i);
874 entry = s->data.plat_entry;
875 io_u_plat = entry->io_u_plat;
877 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
878 io_u_plat_before = entry_before->io_u_plat;
880 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
881 io_sample_ddir(s), (unsigned long long) s->bs);
882 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
883 fprintf(f, "%llu, ", (unsigned long long)
884 hist_sum(j, stride, io_u_plat, io_u_plat_before));
886 fprintf(f, "%llu\n", (unsigned long long)
887 hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
890 flist_del(&entry_before->list);
895 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
899 uint64_t i, nr_samples;
904 s = __get_sample(samples, 0, 0);
905 log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
907 nr_samples = sample_size / __log_entry_sz(log_offset);
909 for (i = 0; i < nr_samples; i++) {
910 s = __get_sample(samples, log_offset, i);
913 fprintf(f, "%lu, %" PRId64 ", %u, %llu, %u\n",
914 (unsigned long) s->time,
916 io_sample_ddir(s), (unsigned long long) s->bs, s->priority_bit);
918 struct io_sample_offset *so = (void *) s;
920 fprintf(f, "%lu, %" PRId64 ", %u, %llu, %llu, %u\n",
921 (unsigned long) s->time,
923 io_sample_ddir(s), (unsigned long long) s->bs,
924 (unsigned long long) so->offset, s->priority_bit);
931 struct iolog_flush_data {
932 struct workqueue_work work;
939 #define GZ_CHUNK 131072
941 static struct iolog_compress *get_new_chunk(unsigned int seq)
943 struct iolog_compress *c;
945 c = malloc(sizeof(*c));
946 INIT_FLIST_HEAD(&c->list);
947 c->buf = malloc(GZ_CHUNK);
953 static void free_chunk(struct iolog_compress *ic)
959 static int z_stream_init(z_stream *stream, int gz_hdr)
963 memset(stream, 0, sizeof(*stream));
964 stream->zalloc = Z_NULL;
965 stream->zfree = Z_NULL;
966 stream->opaque = Z_NULL;
967 stream->next_in = Z_NULL;
970 * zlib magic - add 32 for auto-detection of gz header or not,
971 * if we decide to store files in a gzip friendly format.
976 if (inflateInit2(stream, wbits) != Z_OK)
982 struct inflate_chunk_iter {
991 static void finish_chunk(z_stream *stream, FILE *f,
992 struct inflate_chunk_iter *iter)
996 ret = inflateEnd(stream);
998 log_err("fio: failed to end log inflation seq %d (%d)\n",
1001 flush_samples(f, iter->buf, iter->buf_used);
1004 iter->buf_size = iter->buf_used = 0;
1008 * Iterative chunk inflation. Handles cases where we cross into a new
1009 * sequence, doing flush finish of previous chunk if needed.
1011 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
1012 z_stream *stream, struct inflate_chunk_iter *iter)
1016 dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
1017 (unsigned long) ic->len, ic->seq);
1019 if (ic->seq != iter->seq) {
1021 finish_chunk(stream, f, iter);
1023 z_stream_init(stream, gz_hdr);
1024 iter->seq = ic->seq;
1027 stream->avail_in = ic->len;
1028 stream->next_in = ic->buf;
1030 if (!iter->buf_size) {
1031 iter->buf_size = iter->chunk_sz;
1032 iter->buf = malloc(iter->buf_size);
1035 while (stream->avail_in) {
1036 size_t this_out = iter->buf_size - iter->buf_used;
1039 stream->avail_out = this_out;
1040 stream->next_out = iter->buf + iter->buf_used;
1042 err = inflate(stream, Z_NO_FLUSH);
1044 log_err("fio: failed inflating log: %d\n", err);
1049 iter->buf_used += this_out - stream->avail_out;
1051 if (!stream->avail_out) {
1052 iter->buf_size += iter->chunk_sz;
1053 iter->buf = realloc(iter->buf, iter->buf_size);
1057 if (err == Z_STREAM_END)
1061 ret = (void *) stream->next_in - ic->buf;
1063 dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
1069 * Inflate stored compressed chunks, or write them directly to the log
1070 * file if so instructed.
1072 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1074 struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
1077 while (!flist_empty(&log->chunk_list)) {
1078 struct iolog_compress *ic;
1080 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
1081 flist_del(&ic->list);
1083 if (log->log_gz_store) {
1086 dprint(FD_COMPRESS, "log write chunk size=%lu, "
1087 "seq=%u\n", (unsigned long) ic->len, ic->seq);
1089 ret = fwrite(ic->buf, ic->len, 1, f);
1090 if (ret != 1 || ferror(f)) {
1092 log_err("fio: error writing compressed log\n");
1095 inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
1101 finish_chunk(&stream, f, &iter);
1109 * Open compressed log file and decompress the stored chunks and
1110 * write them to stdout. The chunks are stored sequentially in the
1111 * file, so we iterate over them and do them one-by-one.
1113 int iolog_file_inflate(const char *file)
1115 struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
1116 struct iolog_compress ic;
1124 f = fopen(file, "r");
1130 if (stat(file, &sb) < 0) {
1136 ic.buf = buf = malloc(sb.st_size);
1137 ic.len = sb.st_size;
1140 ret = fread(ic.buf, ic.len, 1, f);
1141 if (ret == 0 && ferror(f)) {
1146 } else if (ferror(f) || (!feof(f) && ret != 1)) {
1147 log_err("fio: short read on reading log\n");
1156 * Each chunk will return Z_STREAM_END. We don't know how many
1157 * chunks are in the file, so we just keep looping and incrementing
1158 * the sequence number until we have consumed the whole compressed
1165 iret = inflate_chunk(&ic, 1, stdout, &stream, &iter);
1178 finish_chunk(&stream, stdout, &iter);
1188 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1193 int iolog_file_inflate(const char *file)
1195 log_err("fio: log inflation not possible without zlib\n");
1201 void flush_log(struct io_log *log, bool do_append)
1207 f = fopen(log->filename, "w");
1209 f = fopen(log->filename, "a");
1211 perror("fopen log");
1215 buf = set_file_buffer(f);
1217 inflate_gz_chunks(log, f);
1219 while (!flist_empty(&log->io_logs)) {
1220 struct io_logs *cur_log;
1222 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1223 flist_del_init(&cur_log->list);
1225 if (log->td && log == log->td->clat_hist_log)
1226 flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1227 log_sample_sz(log, cur_log));
1229 flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1235 clear_file_buffer(buf);
1238 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1240 if (td->flags & TD_F_COMPRESS_LOG)
1244 if (fio_trylock_file(log->filename))
1247 fio_lock_file(log->filename);
1249 if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1250 fio_send_iolog(td, log, log->filename);
1252 flush_log(log, !td->o.per_job_logs);
1254 fio_unlock_file(log->filename);
1259 size_t log_chunk_sizes(struct io_log *log)
1261 struct flist_head *entry;
1264 if (flist_empty(&log->chunk_list))
1268 pthread_mutex_lock(&log->chunk_lock);
1269 flist_for_each(entry, &log->chunk_list) {
1270 struct iolog_compress *c;
1272 c = flist_entry(entry, struct iolog_compress, list);
1275 pthread_mutex_unlock(&log->chunk_lock);
1281 static void iolog_put_deferred(struct io_log *log, void *ptr)
1286 pthread_mutex_lock(&log->deferred_free_lock);
1287 if (log->deferred < IOLOG_MAX_DEFER) {
1288 log->deferred_items[log->deferred] = ptr;
1290 } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
1291 log_err("fio: had to drop log entry free\n");
1292 pthread_mutex_unlock(&log->deferred_free_lock);
1295 static void iolog_free_deferred(struct io_log *log)
1302 pthread_mutex_lock(&log->deferred_free_lock);
1304 for (i = 0; i < log->deferred; i++) {
1305 free(log->deferred_items[i]);
1306 log->deferred_items[i] = NULL;
1310 pthread_mutex_unlock(&log->deferred_free_lock);
1313 static int gz_work(struct iolog_flush_data *data)
1315 struct iolog_compress *c = NULL;
1316 struct flist_head list;
1322 INIT_FLIST_HEAD(&list);
1324 memset(&stream, 0, sizeof(stream));
1325 stream.zalloc = Z_NULL;
1326 stream.zfree = Z_NULL;
1327 stream.opaque = Z_NULL;
1329 ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1331 log_err("fio: failed to init gz stream\n");
1335 seq = ++data->log->chunk_seq;
1337 stream.next_in = (void *) data->samples;
1338 stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1340 dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1341 (unsigned long) stream.avail_in, seq,
1342 data->log->filename);
1345 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1346 (unsigned long) c->len);
1347 c = get_new_chunk(seq);
1348 stream.avail_out = GZ_CHUNK;
1349 stream.next_out = c->buf;
1350 ret = deflate(&stream, Z_NO_FLUSH);
1352 log_err("fio: deflate log (%d)\n", ret);
1357 c->len = GZ_CHUNK - stream.avail_out;
1358 flist_add_tail(&c->list, &list);
1360 } while (stream.avail_in);
1362 stream.next_out = c->buf + c->len;
1363 stream.avail_out = GZ_CHUNK - c->len;
1365 ret = deflate(&stream, Z_FINISH);
1368 * Z_BUF_ERROR is special, it just means we need more
1369 * output space. We'll handle that below. Treat any other
1372 if (ret != Z_BUF_ERROR) {
1373 log_err("fio: deflate log (%d)\n", ret);
1374 flist_del(&c->list);
1381 c->len = GZ_CHUNK - stream.avail_out;
1383 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1385 if (ret != Z_STREAM_END) {
1387 c = get_new_chunk(seq);
1388 stream.avail_out = GZ_CHUNK;
1389 stream.next_out = c->buf;
1390 ret = deflate(&stream, Z_FINISH);
1391 c->len = GZ_CHUNK - stream.avail_out;
1393 flist_add_tail(&c->list, &list);
1394 dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1395 (unsigned long) c->len);
1396 } while (ret != Z_STREAM_END);
1399 dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1401 ret = deflateEnd(&stream);
1403 log_err("fio: deflateEnd %d\n", ret);
1405 iolog_put_deferred(data->log, data->samples);
1407 if (!flist_empty(&list)) {
1408 pthread_mutex_lock(&data->log->chunk_lock);
1409 flist_splice_tail(&list, &data->log->chunk_list);
1410 pthread_mutex_unlock(&data->log->chunk_lock);
1419 while (!flist_empty(&list)) {
1420 c = flist_first_entry(list.next, struct iolog_compress, list);
1421 flist_del(&c->list);
1429 * Invoked from our compress helper thread, when logging would have exceeded
1430 * the specified memory limitation. Compresses the previously stored
1433 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1435 return gz_work(container_of(work, struct iolog_flush_data, work));
1438 static int gz_init_worker(struct submit_worker *sw)
1440 struct thread_data *td = sw->wq->td;
1442 if (!fio_option_is_set(&td->o, log_gz_cpumask))
1445 if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1446 log_err("gz: failed to set CPU affinity\n");
1453 static struct workqueue_ops log_compress_wq_ops = {
1454 .fn = gz_work_async,
1455 .init_worker_fn = gz_init_worker,
1459 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1461 if (!(td->flags & TD_F_COMPRESS_LOG))
1464 workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1468 void iolog_compress_exit(struct thread_data *td)
1470 if (!(td->flags & TD_F_COMPRESS_LOG))
1473 workqueue_exit(&td->log_compress_wq);
1477 * Queue work item to compress the existing log entries. We reset the
1478 * current log to a small size, and reference the existing log in the
1479 * data that we queue for compression. Once compression has been done,
1480 * this old log is freed. If called with finish == true, will not return
1481 * until the log compression has completed, and will flush all previous
1484 static int iolog_flush(struct io_log *log)
1486 struct iolog_flush_data *data;
1488 data = malloc(sizeof(*data));
1495 while (!flist_empty(&log->io_logs)) {
1496 struct io_logs *cur_log;
1498 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1499 flist_del_init(&cur_log->list);
1501 data->samples = cur_log->log;
1502 data->nr_samples = cur_log->nr_samples;
1513 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1515 struct iolog_flush_data *data;
1517 data = smalloc(sizeof(*data));
1523 data->samples = cur_log->log;
1524 data->nr_samples = cur_log->nr_samples;
1527 cur_log->nr_samples = cur_log->max_samples = 0;
1528 cur_log->log = NULL;
1530 workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1532 iolog_free_deferred(log);
1538 static int iolog_flush(struct io_log *log)
1543 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1548 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1553 void iolog_compress_exit(struct thread_data *td)
1559 struct io_logs *iolog_cur_log(struct io_log *log)
1561 if (flist_empty(&log->io_logs))
1564 return flist_last_entry(&log->io_logs, struct io_logs, list);
1567 uint64_t iolog_nr_samples(struct io_log *iolog)
1569 struct flist_head *entry;
1572 flist_for_each(entry, &iolog->io_logs) {
1573 struct io_logs *cur_log;
1575 cur_log = flist_entry(entry, struct io_logs, list);
1576 ret += cur_log->nr_samples;
1582 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1585 return finish_log(td, log, try);
1590 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1594 if (per_unit_log(td->iops_log) != unit_log)
1597 ret = __write_log(td, td->iops_log, try);
1599 td->iops_log = NULL;
1604 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1611 ret = __write_log(td, td->slat_log, try);
1613 td->slat_log = NULL;
1618 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1625 ret = __write_log(td, td->clat_log, try);
1627 td->clat_log = NULL;
1632 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1639 ret = __write_log(td, td->clat_hist_log, try);
1641 td->clat_hist_log = NULL;
1646 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1653 ret = __write_log(td, td->lat_log, try);
1660 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1664 if (per_unit_log(td->bw_log) != unit_log)
1667 ret = __write_log(td, td->bw_log, try);
1680 CLAT_HIST_LOG_MASK = 32,
1687 int (*fn)(struct thread_data *, int, bool);
1690 static struct log_type log_types[] = {
1692 .mask = BW_LOG_MASK,
1693 .fn = write_bandw_log,
1696 .mask = LAT_LOG_MASK,
1697 .fn = write_lat_log,
1700 .mask = SLAT_LOG_MASK,
1701 .fn = write_slat_log,
1704 .mask = CLAT_LOG_MASK,
1705 .fn = write_clat_log,
1708 .mask = IOPS_LOG_MASK,
1709 .fn = write_iops_log,
1712 .mask = CLAT_HIST_LOG_MASK,
1713 .fn = write_clat_hist_log,
1717 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1719 unsigned int log_mask = 0;
1720 unsigned int log_left = ALL_LOG_NR;
1723 old_state = td_bump_runstate(td, TD_FINISHING);
1725 finalize_logs(td, unit_logs);
1728 int prev_log_left = log_left;
1730 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1731 struct log_type *lt = &log_types[i];
1734 if (!(log_mask & lt->mask)) {
1735 ret = lt->fn(td, log_left != 1, unit_logs);
1738 log_mask |= lt->mask;
1743 if (prev_log_left == log_left)
1747 td_restore_runstate(td, old_state);
1750 void fio_writeout_logs(bool unit_logs)
1752 struct thread_data *td;
1756 td_writeout_logs(td, unit_logs);