2 * blktrace support code for fio
11 #include "blktrace_api.h"
12 #include "oslib/linux-dev-lookup.h"
14 #define TRACE_FIFO_SIZE 8192
17 * fifo refill frontend, to avoid reading data in trace sized bites
19 static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
21 char buf[TRACE_FIFO_SIZE];
26 if (total > fifo_room(fifo))
27 total = fifo_room(fifo);
29 ret = read(fd, buf, total);
31 td_verror(td, errno, "read blktrace file");
36 ret = fifo_put(fifo, buf, ret);
38 dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
43 * Retrieve 'len' bytes from the fifo, refilling if necessary.
45 static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
46 void *buf, unsigned int len)
48 if (fifo_len(fifo) < len) {
49 int ret = refill_fifo(td, fifo, fd);
55 return fifo_get(fifo, buf, len);
59 * Just discard the pdu by seeking past it.
61 static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
62 struct blk_io_trace *t)
67 dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
68 return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
72 * Check if this is a blktrace binary data file. We read a single trace
73 * into memory and check for the magic signature.
75 bool is_blktrace(const char *filename, int *need_swap)
77 struct blk_io_trace t;
80 fd = open(filename, O_RDONLY);
84 ret = read(fd, &t, sizeof(t));
88 perror("read blktrace");
90 } else if (ret != sizeof(t)) {
91 log_err("fio: short read on blktrace file\n");
95 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
101 * Maybe it needs to be endian swapped...
103 t.magic = fio_swap32(t.magic);
104 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
112 #define FMINORBITS 20
113 #define FMINORMASK ((1U << FMINORBITS) - 1)
114 #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
115 #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
117 static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
119 struct io_piece *ipo;
121 ipo = calloc(1, sizeof(*ipo));
124 ipo->ddir = DDIR_INVAL;
125 ipo->fileno = fileno;
126 ipo->file_action = action;
127 flist_add_tail(&ipo->list, &td->io_log_list);
130 static int trace_add_file(struct thread_data *td, __u32 device)
132 static unsigned int last_maj, last_min, last_fileno;
133 unsigned int maj = FMAJOR(device);
134 unsigned int min = FMINOR(device);
139 if (last_maj == maj && last_min == min)
146 * check for this file in our list
148 for_each_file(td, f, i)
149 if (f->major == maj && f->minor == min) {
150 last_fileno = f->fileno;
155 if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
158 if (td->o.replay_redirect)
159 dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
160 " with: %s\n", maj, min,
161 td->o.replay_redirect);
163 dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
165 dprint(FD_BLKTRACE, "add devices %s\n", dev);
166 fileno = add_file_exclusive(td, dev);
168 td->files[fileno]->major = maj;
169 td->files[fileno]->minor = min;
170 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
171 last_fileno = fileno;
177 static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
179 if (!o->replay_align)
182 t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
186 * Store blk_io_trace data in an ipo for later retrieval.
188 static void store_ipo(struct thread_data *td, unsigned long long offset,
189 unsigned int bytes, int rw, unsigned long long ttime,
192 struct io_piece *ipo;
194 ipo = calloc(1, sizeof(*ipo));
197 ipo->offset = offset * 512;
198 if (td->o.replay_scale)
199 ipo->offset = ipo->offset / td->o.replay_scale;
200 ipo_bytes_align(td->o.replay_align, ipo);
202 ipo->delay = ttime / 1000;
204 ipo->ddir = DDIR_WRITE;
206 ipo->ddir = DDIR_READ;
207 ipo->fileno = fileno;
209 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
210 ipo->ddir, ipo->offset,
211 ipo->len, ipo->delay);
212 queue_io_piece(td, ipo);
215 static void handle_trace_notify(struct blk_io_trace *t)
219 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
222 case BLK_TN_TIMESTAMP:
223 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
229 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
234 static void handle_trace_discard(struct thread_data *td,
235 struct blk_io_trace *t,
236 unsigned long long ttime,
237 unsigned long *ios, unsigned int *bs)
239 struct io_piece *ipo;
242 if (td->o.replay_skip & (1u << DDIR_TRIM))
245 ipo = calloc(1, sizeof(*ipo));
247 fileno = trace_add_file(td, t->device);
250 if (t->bytes > bs[DDIR_TRIM])
251 bs[DDIR_TRIM] = t->bytes;
253 td->o.size += t->bytes;
255 INIT_FLIST_HEAD(&ipo->list);
257 ipo->offset = t->sector * 512;
258 if (td->o.replay_scale)
259 ipo->offset = ipo->offset / td->o.replay_scale;
260 ipo_bytes_align(td->o.replay_align, ipo);
262 ipo->delay = ttime / 1000;
263 ipo->ddir = DDIR_TRIM;
264 ipo->fileno = fileno;
266 dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
267 ipo->offset, ipo->len,
269 queue_io_piece(td, ipo);
272 static void dump_trace(struct blk_io_trace *t)
274 log_err("blktrace: ignoring zero byte trace: action=%x\n", t->action);
277 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
278 unsigned long long ttime, unsigned long *ios,
284 fileno = trace_add_file(td, t->device);
286 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
289 if (td->o.replay_skip & (1u << DDIR_WRITE))
292 if (td->o.replay_skip & (1u << DDIR_READ))
297 if (!fio_did_warn(FIO_WARN_BTRACE_ZERO))
302 if (t->bytes > bs[rw])
306 td->o.size += t->bytes;
307 store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
310 static void handle_trace_flush(struct thread_data *td, struct blk_io_trace *t,
311 unsigned long long ttime, unsigned long *ios)
313 struct io_piece *ipo;
316 if (td->o.replay_skip & (1u << DDIR_SYNC))
319 ipo = calloc(1, sizeof(*ipo));
321 fileno = trace_add_file(td, t->device);
323 ipo->delay = ttime / 1000;
324 ipo->ddir = DDIR_SYNC;
325 ipo->fileno = fileno;
328 dprint(FD_BLKTRACE, "store flush delay=%lu\n", ipo->delay);
329 queue_io_piece(td, ipo);
333 * We only care for queue traces, most of the others are side effects
334 * due to internal workings of the block layer.
336 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
337 unsigned long *ios, unsigned int *bs)
339 static unsigned long long last_ttime;
340 unsigned long long delay = 0;
342 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
345 if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
346 if (!last_ttime || td->o.no_stall)
348 else if (td->o.replay_time_scale == 100)
349 delay = t->time - last_ttime;
351 double tmp = t->time - last_ttime;
354 scale = (double) 100.0 / (double) td->o.replay_time_scale;
358 last_ttime = t->time;
361 t_bytes_align(&td->o, t);
363 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
364 handle_trace_notify(t);
365 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
366 handle_trace_discard(td, t, delay, ios, bs);
367 else if (t->action & BLK_TC_ACT(BLK_TC_FLUSH))
368 handle_trace_flush(td, t, delay, ios);
370 handle_trace_fs(td, t, delay, ios, bs);
373 static void byteswap_trace(struct blk_io_trace *t)
375 t->magic = fio_swap32(t->magic);
376 t->sequence = fio_swap32(t->sequence);
377 t->time = fio_swap64(t->time);
378 t->sector = fio_swap64(t->sector);
379 t->bytes = fio_swap32(t->bytes);
380 t->action = fio_swap32(t->action);
381 t->pid = fio_swap32(t->pid);
382 t->device = fio_swap32(t->device);
383 t->cpu = fio_swap32(t->cpu);
384 t->error = fio_swap16(t->error);
385 t->pdu_len = fio_swap16(t->pdu_len);
388 static bool t_is_write(struct blk_io_trace *t)
390 return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
393 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
395 if (t->action & BLK_TC_ACT(BLK_TC_READ))
397 else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
399 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
405 static void depth_inc(struct blk_io_trace *t, int *depth)
409 ddir = t_get_ddir(t);
410 if (ddir != DDIR_INVAL)
414 static void depth_dec(struct blk_io_trace *t, int *depth)
418 ddir = t_get_ddir(t);
419 if (ddir != DDIR_INVAL)
423 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
425 enum fio_ddir ddir = DDIR_INVAL;
427 ddir = t_get_ddir(t);
428 if (ddir != DDIR_INVAL) {
429 depth[ddir] = max(depth[ddir], this_depth[ddir]);
430 this_depth[ddir] = 0;
435 * Load a blktrace file by reading all the blk_io_trace entries, and storing
436 * them as io_pieces like the fio text version would do.
438 bool load_blktrace(struct thread_data *td, const char *filename, int need_swap)
440 struct blk_io_trace t;
441 unsigned long ios[DDIR_RWDIR_SYNC_CNT] = { };
442 unsigned int rw_bs[DDIR_RWDIR_CNT] = { };
443 unsigned long skipped_writes;
445 int fd, i, old_state, max_depth;
447 int this_depth[DDIR_RWDIR_CNT] = { };
448 int depth[DDIR_RWDIR_CNT] = { };
450 fd = open(filename, O_RDONLY);
452 td_verror(td, errno, "open blktrace file");
456 fifo = fifo_alloc(TRACE_FIFO_SIZE);
458 old_state = td_bump_runstate(td, TD_SETTING_UP);
463 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
469 else if (ret < (int) sizeof(t)) {
470 log_err("fio: short fifo get\n");
477 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
478 log_err("fio: bad magic in blktrace data: %x\n",
482 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
483 log_err("fio: bad blktrace version %d\n",
487 ret = discard_pdu(td, fifo, fd, &t);
489 td_verror(td, ret, "blktrace lseek");
491 } else if (t.pdu_len != ret) {
492 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
495 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
496 if ((t.action & 0xffff) == __BLK_TA_QUEUE)
497 depth_inc(&t, this_depth);
498 else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
499 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
500 depth_dec(&t, this_depth);
501 else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
502 depth_end(&t, this_depth, depth);
504 if (t_is_write(&t) && read_only) {
510 handle_trace(td, &t, ios, rw_bs);
513 for_each_file(td, f, i)
514 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
519 td_restore_runstate(td, old_state);
521 if (!td->files_index) {
522 log_err("fio: did not find replay device(s)\n");
527 * For stacked devices, we don't always get a COMPLETE event so
528 * the depth grows to insane values. Limit it to something sane(r).
531 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
534 else if (!depth[i] && ios[i])
536 max_depth = max(depth[i], max_depth);
540 log_err("fio: %s skips replay of %lu writes due to read-only\n",
541 td->o.name, skipped_writes);
543 if (!ios[DDIR_READ] && !ios[DDIR_WRITE] && !ios[DDIR_TRIM] &&
545 log_err("fio: found no ios in blktrace data\n");
550 if (ios[DDIR_READ]) {
551 td->o.td_ddir |= TD_DDIR_READ;
552 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
554 if (ios[DDIR_WRITE]) {
555 td->o.td_ddir |= TD_DDIR_WRITE;
556 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
558 if (ios[DDIR_TRIM]) {
559 td->o.td_ddir |= TD_DDIR_TRIM;
560 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
564 * We need to do direct/raw ios to the device, to avoid getting
565 * read-ahead in our way. But only do so if the minimum block size
566 * is a multiple of 4k, otherwise we don't know if it's safe to do so.
568 if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
572 * If depth wasn't manually set, use probed depth
574 if (!fio_option_is_set(&td->o, iodepth))
575 td->o.iodepth = td->o.iodepth_low = max_depth;
584 static int init_merge_param_list(fio_fp64_t *vals, struct blktrace_cursor *bcs,
585 int nr_logs, int def, size_t off)
589 while (len < FIO_IO_U_LIST_MAX_LEN && vals[len].u.f != 0.0)
592 if (len && len != nr_logs)
595 for (i = 0; i < nr_logs; i++) {
596 int *val = (int *)((char *)&bcs[i] + off);
599 *val = (int)vals[i].u.f;
606 static int find_earliest_io(struct blktrace_cursor *bcs, int nr_logs)
608 __u64 time = ~(__u64)0;
611 for (i = 0; i < nr_logs; i++) {
612 if (bcs[i].t.time < time) {
613 time = bcs[i].t.time;
621 static void merge_finish_file(struct blktrace_cursor *bcs, int i, int *nr_logs)
624 if (bcs[i].iter < bcs[i].nr_iter) {
625 lseek(bcs[i].fd, 0, SEEK_SET);
632 fifo_free(bcs[i].fifo);
635 /* keep active files contiguous */
636 memmove(&bcs[i], &bcs[*nr_logs], sizeof(bcs[i]));
639 static int read_trace(struct thread_data *td, struct blktrace_cursor *bc)
642 struct blk_io_trace *t = &bc->t;
645 /* read an io trace */
646 ret = trace_fifo_get(td, bc->fifo, bc->fd, t, sizeof(*t));
651 bc->length = bc->t.time;
653 } else if (ret < (int) sizeof(*t)) {
654 log_err("fio: short fifo get\n");
661 /* skip over actions that fio does not care about */
662 if ((t->action & 0xffff) != __BLK_TA_QUEUE ||
663 t_get_ddir(t) == DDIR_INVAL) {
664 ret = discard_pdu(td, bc->fifo, bc->fd, t);
666 td_verror(td, ret, "blktrace lseek");
668 } else if (t->pdu_len != ret) {
669 log_err("fio: discarded %d of %d\n", ret,
676 t->time = (t->time + bc->iter * bc->length) * bc->scalar / 100;
681 static int write_trace(FILE *fp, struct blk_io_trace *t)
683 /* pdu is not used so just write out only the io trace */
685 return fwrite((void *)t, sizeof(*t), 1, fp);
688 int merge_blktrace_iologs(struct thread_data *td)
690 int nr_logs = get_max_str_idx(td->o.read_iolog_file);
691 struct blktrace_cursor *bcs = malloc(sizeof(struct blktrace_cursor) *
693 struct blktrace_cursor *bc;
695 char *str, *ptr, *name, *merge_buf;
698 ret = init_merge_param_list(td->o.merge_blktrace_scalars, bcs, nr_logs,
699 100, offsetof(struct blktrace_cursor,
702 log_err("fio: merge_blktrace_scalars(%d) != nr_logs(%d)\n",
707 ret = init_merge_param_list(td->o.merge_blktrace_iters, bcs, nr_logs,
708 1, offsetof(struct blktrace_cursor,
711 log_err("fio: merge_blktrace_iters(%d) != nr_logs(%d)\n",
716 /* setup output file */
717 merge_fp = fopen(td->o.merge_blktrace_file, "w");
718 merge_buf = malloc(128 * 1024);
719 ret = setvbuf(merge_fp, merge_buf, _IOFBF, 128 * 1024);
723 /* setup input files */
724 str = ptr = strdup(td->o.read_iolog_file);
726 for (i = 0; (name = get_next_str(&ptr)) != NULL; i++) {
727 bcs[i].fd = open(name, O_RDONLY);
729 log_err("fio: could not open file: %s\n", name);
733 bcs[i].fifo = fifo_alloc(TRACE_FIFO_SIZE);
736 if (!is_blktrace(name, &bcs[i].swap)) {
737 log_err("fio: file is not a blktrace: %s\n", name);
741 ret = read_trace(td, &bcs[i]);
745 merge_finish_file(bcs, i, &nr_logs);
753 i = find_earliest_io(bcs, nr_logs);
755 /* skip over the pdu */
756 ret = discard_pdu(td, bc->fifo, bc->fd, &bc->t);
758 td_verror(td, ret, "blktrace lseek");
760 } else if (bc->t.pdu_len != ret) {
761 log_err("fio: discarded %d of %d\n", ret,
766 ret = write_trace(merge_fp, &bc->t);
767 ret = read_trace(td, bc);
771 merge_finish_file(bcs, i, &nr_logs);
774 /* set iolog file to read from the newly merged file */
775 td->o.read_iolog_file = td->o.merge_blktrace_file;
780 for (i = 0; i < nr_logs; i++) {
781 fifo_free(bcs[i].fifo);