2 * blktrace support code for fio
13 #include "blktrace_api.h"
14 #include "oslib/linux-dev-lookup.h"
16 #define TRACE_FIFO_SIZE 8192
19 * fifo refill frontend, to avoid reading data in trace sized bites
21 static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
23 char buf[TRACE_FIFO_SIZE];
28 if (total > fifo_room(fifo))
29 total = fifo_room(fifo);
31 ret = read(fd, buf, total);
33 td_verror(td, errno, "read blktrace file");
38 ret = fifo_put(fifo, buf, ret);
40 dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
45 * Retrieve 'len' bytes from the fifo, refilling if necessary.
47 static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
48 void *buf, unsigned int len)
50 if (fifo_len(fifo) < len) {
51 int ret = refill_fifo(td, fifo, fd);
57 return fifo_get(fifo, buf, len);
61 * Just discard the pdu by seeking past it.
63 static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
64 struct blk_io_trace *t)
69 dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
70 return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
74 * Check if this is a blktrace binary data file. We read a single trace
75 * into memory and check for the magic signature.
77 int is_blktrace(const char *filename, int *need_swap)
79 struct blk_io_trace t;
82 fd = open(filename, O_RDONLY);
86 ret = read(fd, &t, sizeof(t));
90 perror("read blktrace");
92 } else if (ret != sizeof(t)) {
93 log_err("fio: short read on blktrace file\n");
97 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
103 * Maybe it needs to be endian swapped...
105 t.magic = fio_swap32(t.magic);
106 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
114 #define FMINORBITS 20
115 #define FMINORMASK ((1U << FMINORBITS) - 1)
116 #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
117 #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
119 static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
121 struct io_piece *ipo;
123 ipo = calloc(1, sizeof(*ipo));
126 ipo->ddir = DDIR_INVAL;
127 ipo->fileno = fileno;
128 ipo->file_action = action;
129 flist_add_tail(&ipo->list, &td->io_log_list);
132 static int get_dev_blocksize(const char *dev, unsigned int *bs)
136 fd = open(dev, O_RDONLY);
140 if (ioctl(fd, BLKSSZGET, bs) < 0) {
149 static int trace_add_file(struct thread_data *td, __u32 device,
152 static unsigned int last_maj, last_min, last_fileno, last_bs;
153 unsigned int maj = FMAJOR(device);
154 unsigned int min = FMINOR(device);
159 if (last_maj == maj && last_min == min) {
168 * check for this file in our list
170 for_each_file(td, f, i) {
171 if (f->major == maj && f->minor == min) {
172 last_fileno = f->fileno;
179 if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
180 unsigned int this_bs;
183 if (td->o.replay_redirect)
184 dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
185 " with: %s\n", maj, min,
186 td->o.replay_redirect);
188 dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
190 dprint(FD_BLKTRACE, "add devices %s\n", dev);
191 fileno = add_file_exclusive(td, dev);
193 if (get_dev_blocksize(dev, &this_bs))
197 td->files[fileno]->major = maj;
198 td->files[fileno]->minor = min;
199 td->files[fileno]->bs = this_bs;
200 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
202 last_fileno = fileno;
211 static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
213 if (!o->replay_align)
216 t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
219 static void ipo_bytes_align(struct thread_options *o, struct io_piece *ipo)
221 if (!o->replay_align)
224 ipo->offset &= ~(o->replay_align - 1);
229 * Store blk_io_trace data in an ipo for later retrieval.
231 static void store_ipo(struct thread_data *td, unsigned long long offset,
232 unsigned int bytes, int rw, unsigned long long ttime,
233 int fileno, unsigned int bs)
235 struct io_piece *ipo = malloc(sizeof(*ipo));
239 ipo->offset = offset * bs;
240 if (td->o.replay_scale)
241 ipo->offset = ipo->offset / td->o.replay_scale;
242 ipo_bytes_align(&td->o, ipo);
244 ipo->delay = ttime / 1000;
246 ipo->ddir = DDIR_WRITE;
248 ipo->ddir = DDIR_READ;
249 ipo->fileno = fileno;
251 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
252 ipo->ddir, ipo->offset,
253 ipo->len, ipo->delay);
254 queue_io_piece(td, ipo);
257 static void handle_trace_notify(struct blk_io_trace *t)
261 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
264 case BLK_TN_TIMESTAMP:
265 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
271 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
276 static void handle_trace_discard(struct thread_data *td,
277 struct blk_io_trace *t,
278 unsigned long long ttime,
279 unsigned long *ios, unsigned int *rw_bs)
281 struct io_piece *ipo = malloc(sizeof(*ipo));
286 fileno = trace_add_file(td, t->device, &bs);
289 if (t->bytes > rw_bs[DDIR_TRIM])
290 rw_bs[DDIR_TRIM] = t->bytes;
292 td->o.size += t->bytes;
294 memset(ipo, 0, sizeof(*ipo));
295 INIT_FLIST_HEAD(&ipo->list);
297 ipo->offset = t->sector * bs;
298 if (td->o.replay_scale)
299 ipo->offset = ipo->offset / td->o.replay_scale;
300 ipo_bytes_align(&td->o, ipo);
302 ipo->delay = ttime / 1000;
303 ipo->ddir = DDIR_TRIM;
304 ipo->fileno = fileno;
306 dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
307 ipo->offset, ipo->len,
309 queue_io_piece(td, ipo);
312 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
313 unsigned long long ttime, unsigned long *ios,
320 fileno = trace_add_file(td, t->device, &bs);
322 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
324 if (t->bytes > rw_bs[rw])
325 rw_bs[rw] = t->bytes;
328 td->o.size += t->bytes;
329 store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
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) {
347 last_ttime = t->time;
350 delay = t->time - last_ttime;
351 last_ttime = t->time;
355 t_bytes_align(&td->o, t);
357 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
358 handle_trace_notify(t);
359 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
360 handle_trace_discard(td, t, delay, ios, bs);
362 handle_trace_fs(td, t, delay, ios, bs);
365 static void byteswap_trace(struct blk_io_trace *t)
367 t->magic = fio_swap32(t->magic);
368 t->sequence = fio_swap32(t->sequence);
369 t->time = fio_swap64(t->time);
370 t->sector = fio_swap64(t->sector);
371 t->bytes = fio_swap32(t->bytes);
372 t->action = fio_swap32(t->action);
373 t->pid = fio_swap32(t->pid);
374 t->device = fio_swap32(t->device);
375 t->cpu = fio_swap32(t->cpu);
376 t->error = fio_swap16(t->error);
377 t->pdu_len = fio_swap16(t->pdu_len);
380 static int t_is_write(struct blk_io_trace *t)
382 return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
385 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
387 if (t->action & BLK_TC_ACT(BLK_TC_READ))
389 else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
391 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
397 static void depth_inc(struct blk_io_trace *t, int *depth)
401 ddir = t_get_ddir(t);
402 if (ddir != DDIR_INVAL)
406 static void depth_dec(struct blk_io_trace *t, int *depth)
410 ddir = t_get_ddir(t);
411 if (ddir != DDIR_INVAL)
415 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
417 enum fio_ddir ddir = DDIR_INVAL;
419 ddir = t_get_ddir(t);
420 if (ddir != DDIR_INVAL) {
421 depth[ddir] = max(depth[ddir], this_depth[ddir]);
422 this_depth[ddir] = 0;
427 * Load a blktrace file by reading all the blk_io_trace entries, and storing
428 * them as io_pieces like the fio text version would do.
430 int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
432 struct blk_io_trace t;
433 unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
434 unsigned int rw_bs[DDIR_RWDIR_CNT];
436 int fd, i, old_state;
438 int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
440 fd = open(filename, O_RDONLY);
442 td_verror(td, errno, "open blktrace file");
446 fifo = fifo_alloc(TRACE_FIFO_SIZE);
448 old_state = td_bump_runstate(td, TD_SETTING_UP);
452 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
461 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
467 else if (ret < (int) sizeof(t)) {
468 log_err("fio: short fifo get\n");
475 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
476 log_err("fio: bad magic in blktrace data: %x\n",
480 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
481 log_err("fio: bad blktrace version %d\n",
485 ret = discard_pdu(td, fifo, fd, &t);
487 td_verror(td, ret, "blktrace lseek");
489 } else if (t.pdu_len != ret) {
490 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
493 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
494 if ((t.action & 0xffff) == __BLK_TA_QUEUE)
495 depth_inc(&t, this_depth);
496 else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
497 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
498 depth_dec(&t, this_depth);
499 else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
500 depth_end(&t, this_depth, depth);
502 if (t_is_write(&t) && read_only) {
508 handle_trace(td, &t, ios, rw_bs);
511 for (i = 0; i < td->files_index; i++) {
513 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]) {
544 log_err("fio: found no ios in blktrace data\n");
546 } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
547 td->o.td_ddir = TD_DDIR_READ;
548 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
549 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
550 td->o.td_ddir = TD_DDIR_WRITE;
551 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
553 td->o.td_ddir = TD_DDIR_RW;
554 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
555 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
556 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
560 * We need to do direct/raw ios to the device, to avoid getting
561 * read-ahead in our way. But only do so if the minimum block size
562 * is a multiple of 4k, otherwise we don't know if it's safe to do so.
564 if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
568 * If depth wasn't manually set, use probed depth
570 if (!fio_option_is_set(&td->o, iodepth))
571 td->o.iodepth = td->o.iodepth_low = max_depth;