2 * blktrace support code for fio
12 #include "blktrace_api.h"
13 #include "oslib/linux-dev-lookup.h"
15 #define TRACE_FIFO_SIZE 8192
18 * fifo refill frontend, to avoid reading data in trace sized bites
20 static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
22 char buf[TRACE_FIFO_SIZE];
27 if (total > fifo_room(fifo))
28 total = fifo_room(fifo);
30 ret = read(fd, buf, total);
32 td_verror(td, errno, "read blktrace file");
37 ret = fifo_put(fifo, buf, ret);
39 dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
44 * Retrieve 'len' bytes from the fifo, refilling if necessary.
46 static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
47 void *buf, unsigned int len)
49 if (fifo_len(fifo) < len) {
50 int ret = refill_fifo(td, fifo, fd);
56 return fifo_get(fifo, buf, len);
60 * Just discard the pdu by seeking past it.
62 static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
63 struct blk_io_trace *t)
68 dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
69 return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
73 * Check if this is a blktrace binary data file. We read a single trace
74 * into memory and check for the magic signature.
76 int is_blktrace(const char *filename, int *need_swap)
78 struct blk_io_trace t;
81 fd = open(filename, O_RDONLY);
85 ret = read(fd, &t, sizeof(t));
89 perror("read blktrace");
91 } else if (ret != sizeof(t)) {
92 log_err("fio: short read on blktrace file\n");
96 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
102 * Maybe it needs to be endian swapped...
104 t.magic = fio_swap32(t.magic);
105 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
113 #define FMINORBITS 20
114 #define FMINORMASK ((1U << FMINORBITS) - 1)
115 #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
116 #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
118 static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
120 struct io_piece *ipo;
122 ipo = calloc(1, sizeof(*ipo));
125 ipo->ddir = DDIR_INVAL;
126 ipo->fileno = fileno;
127 ipo->file_action = action;
128 flist_add_tail(&ipo->list, &td->io_log_list);
131 static int get_dev_blocksize(const char *dev, unsigned int *bs)
135 fd = open(dev, O_RDONLY);
139 if (ioctl(fd, BLKSSZGET, bs) < 0) {
148 static int trace_add_file(struct thread_data *td, __u32 device,
151 static unsigned int last_maj, last_min, last_fileno, last_bs;
152 unsigned int maj = FMAJOR(device);
153 unsigned int min = FMINOR(device);
158 if (last_maj == maj && last_min == min) {
167 * check for this file in our list
169 for_each_file(td, f, i) {
170 if (f->major == maj && f->minor == min) {
171 last_fileno = f->fileno;
178 if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
179 unsigned int this_bs;
182 if (td->o.replay_redirect)
183 dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
184 " with: %s\n", maj, min,
185 td->o.replay_redirect);
187 dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
189 dprint(FD_BLKTRACE, "add devices %s\n", dev);
190 fileno = add_file_exclusive(td, dev);
192 if (get_dev_blocksize(dev, &this_bs))
196 td->files[fileno]->major = maj;
197 td->files[fileno]->minor = min;
198 td->files[fileno]->bs = this_bs;
199 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
201 last_fileno = fileno;
210 static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
212 if (!o->replay_align)
215 t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
219 * Store blk_io_trace data in an ipo for later retrieval.
221 static void store_ipo(struct thread_data *td, unsigned long long offset,
222 unsigned int bytes, int rw, unsigned long long ttime,
223 int fileno, unsigned int bs)
225 struct io_piece *ipo = malloc(sizeof(*ipo));
229 ipo->offset = offset * bs;
230 if (td->o.replay_scale)
231 ipo->offset = ipo->offset / td->o.replay_scale;
232 ipo_bytes_align(td->o.replay_align, ipo);
234 ipo->delay = ttime / 1000;
236 ipo->ddir = DDIR_WRITE;
238 ipo->ddir = DDIR_READ;
239 ipo->fileno = fileno;
241 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
242 ipo->ddir, ipo->offset,
243 ipo->len, ipo->delay);
244 queue_io_piece(td, ipo);
247 static void handle_trace_notify(struct blk_io_trace *t)
251 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
254 case BLK_TN_TIMESTAMP:
255 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
261 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
266 static void handle_trace_discard(struct thread_data *td,
267 struct blk_io_trace *t,
268 unsigned long long ttime,
269 unsigned long *ios, unsigned int *rw_bs)
271 struct io_piece *ipo = malloc(sizeof(*ipo));
276 fileno = trace_add_file(td, t->device, &bs);
279 if (t->bytes > rw_bs[DDIR_TRIM])
280 rw_bs[DDIR_TRIM] = t->bytes;
282 td->o.size += t->bytes;
284 memset(ipo, 0, sizeof(*ipo));
285 INIT_FLIST_HEAD(&ipo->list);
287 ipo->offset = t->sector * bs;
288 if (td->o.replay_scale)
289 ipo->offset = ipo->offset / td->o.replay_scale;
290 ipo_bytes_align(td->o.replay_align, ipo);
292 ipo->delay = ttime / 1000;
293 ipo->ddir = DDIR_TRIM;
294 ipo->fileno = fileno;
296 dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
297 ipo->offset, ipo->len,
299 queue_io_piece(td, ipo);
302 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
303 unsigned long long ttime, unsigned long *ios,
310 fileno = trace_add_file(td, t->device, &bs);
312 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
314 if (t->bytes > rw_bs[rw])
315 rw_bs[rw] = t->bytes;
318 td->o.size += t->bytes;
319 store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
323 * We only care for queue traces, most of the others are side effects
324 * due to internal workings of the block layer.
326 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
327 unsigned long *ios, unsigned int *bs)
329 static unsigned long long last_ttime;
330 unsigned long long delay = 0;
332 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
335 if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
336 if (!last_ttime || td->o.no_stall) {
337 last_ttime = t->time;
340 delay = t->time - last_ttime;
341 last_ttime = t->time;
345 t_bytes_align(&td->o, t);
347 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
348 handle_trace_notify(t);
349 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
350 handle_trace_discard(td, t, delay, ios, bs);
352 handle_trace_fs(td, t, delay, ios, bs);
355 static void byteswap_trace(struct blk_io_trace *t)
357 t->magic = fio_swap32(t->magic);
358 t->sequence = fio_swap32(t->sequence);
359 t->time = fio_swap64(t->time);
360 t->sector = fio_swap64(t->sector);
361 t->bytes = fio_swap32(t->bytes);
362 t->action = fio_swap32(t->action);
363 t->pid = fio_swap32(t->pid);
364 t->device = fio_swap32(t->device);
365 t->cpu = fio_swap32(t->cpu);
366 t->error = fio_swap16(t->error);
367 t->pdu_len = fio_swap16(t->pdu_len);
370 static int t_is_write(struct blk_io_trace *t)
372 return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
375 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
377 if (t->action & BLK_TC_ACT(BLK_TC_READ))
379 else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
381 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
387 static void depth_inc(struct blk_io_trace *t, int *depth)
391 ddir = t_get_ddir(t);
392 if (ddir != DDIR_INVAL)
396 static void depth_dec(struct blk_io_trace *t, int *depth)
400 ddir = t_get_ddir(t);
401 if (ddir != DDIR_INVAL)
405 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
407 enum fio_ddir ddir = DDIR_INVAL;
409 ddir = t_get_ddir(t);
410 if (ddir != DDIR_INVAL) {
411 depth[ddir] = max(depth[ddir], this_depth[ddir]);
412 this_depth[ddir] = 0;
417 * Load a blktrace file by reading all the blk_io_trace entries, and storing
418 * them as io_pieces like the fio text version would do.
420 int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
422 struct blk_io_trace t;
423 unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
424 unsigned int rw_bs[DDIR_RWDIR_CNT];
426 int fd, i, old_state;
428 int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
430 fd = open(filename, O_RDONLY);
432 td_verror(td, errno, "open blktrace file");
436 fifo = fifo_alloc(TRACE_FIFO_SIZE);
438 old_state = td_bump_runstate(td, TD_SETTING_UP);
442 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
451 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
457 else if (ret < (int) sizeof(t)) {
458 log_err("fio: short fifo get\n");
465 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
466 log_err("fio: bad magic in blktrace data: %x\n",
470 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
471 log_err("fio: bad blktrace version %d\n",
475 ret = discard_pdu(td, fifo, fd, &t);
477 td_verror(td, ret, "blktrace lseek");
479 } else if (t.pdu_len != ret) {
480 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
483 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
484 if ((t.action & 0xffff) == __BLK_TA_QUEUE)
485 depth_inc(&t, this_depth);
486 else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
487 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
488 depth_dec(&t, this_depth);
489 else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
490 depth_end(&t, this_depth, depth);
492 if (t_is_write(&t) && read_only) {
498 handle_trace(td, &t, ios, rw_bs);
501 for_each_file(td, f, i)
502 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
507 td_restore_runstate(td, old_state);
509 if (!td->files_index) {
510 log_err("fio: did not find replay device(s)\n");
515 * For stacked devices, we don't always get a COMPLETE event so
516 * the depth grows to insane values. Limit it to something sane(r).
519 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
522 else if (!depth[i] && ios[i])
524 max_depth = max(depth[i], max_depth);
528 log_err("fio: %s skips replay of %lu writes due to read-only\n",
529 td->o.name, skipped_writes);
531 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
532 log_err("fio: found no ios in blktrace data\n");
534 } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
535 td->o.td_ddir = TD_DDIR_READ;
536 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
537 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
538 td->o.td_ddir = TD_DDIR_WRITE;
539 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
541 td->o.td_ddir = TD_DDIR_RW;
542 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
543 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
544 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
548 * We need to do direct/raw ios to the device, to avoid getting
549 * read-ahead in our way. But only do so if the minimum block size
550 * is a multiple of 4k, otherwise we don't know if it's safe to do so.
552 if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
556 * If depth wasn't manually set, use probed depth
558 if (!fio_option_is_set(&td->o, iodepth))
559 td->o.iodepth = td->o.iodepth_low = max_depth;