2 * blktrace support code for fio
11 #include "blktrace_api.h"
13 #define TRACE_FIFO_SIZE 8192
16 * fifo refill frontend, to avoid reading data in trace sized bites
18 static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
20 char buf[TRACE_FIFO_SIZE];
25 if (total > fifo_room(fifo))
26 total = fifo_room(fifo);
28 ret = read(fd, buf, total);
30 td_verror(td, errno, "read blktrace file");
35 ret = fifo_put(fifo, buf, ret);
37 dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
42 * Retrieve 'len' bytes from the fifo, refilling if necessary.
44 static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
45 void *buf, unsigned int len)
47 if (fifo_len(fifo) < len) {
48 int ret = refill_fifo(td, fifo, fd);
54 return fifo_get(fifo, buf, len);
58 * Just discard the pdu by seeking past it.
60 static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
61 struct blk_io_trace *t)
66 dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
67 return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
71 * Check if this is a blktrace binary data file. We read a single trace
72 * into memory and check for the magic signature.
74 int is_blktrace(const char *filename, int *need_swap)
76 struct blk_io_trace t;
79 fd = open(filename, O_RDONLY);
83 ret = read(fd, &t, sizeof(t));
87 perror("read blktrace");
89 } else if (ret != sizeof(t)) {
90 log_err("fio: short read on blktrace file\n");
94 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
100 * Maybe it needs to be endian swapped...
102 t.magic = fio_swap32(t.magic);
103 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
111 static int lookup_device(struct thread_data *td, char *path, unsigned int maj,
123 while ((dir = readdir(D)) != NULL) {
126 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
129 sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name);
130 if (lstat(full_path, &st) == -1) {
135 if (S_ISDIR(st.st_mode)) {
136 found = lookup_device(td, full_path, maj, min);
138 strcpy(path, full_path);
143 if (!S_ISBLK(st.st_mode))
147 * If replay_redirect is set then always return this device
148 * upon lookup which overrides the device lookup based on
149 * major minor in the actual blktrace
151 if (td->o.replay_redirect) {
152 dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
153 " with: %s", maj, min,
154 td->o.replay_redirect);
155 strcpy(path, td->o.replay_redirect);
160 if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
161 dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
162 strcpy(path, full_path);
172 #define FMINORBITS 20
173 #define FMINORMASK ((1U << FMINORBITS) - 1)
174 #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
175 #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
177 static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
179 struct io_piece *ipo;
181 ipo = calloc(1, sizeof(*ipo));
184 ipo->ddir = DDIR_INVAL;
185 ipo->fileno = fileno;
186 ipo->file_action = action;
187 flist_add_tail(&ipo->list, &td->io_log_list);
190 static int trace_add_file(struct thread_data *td, __u32 device)
192 static unsigned int last_maj, last_min, last_fileno;
193 unsigned int maj = FMAJOR(device);
194 unsigned int min = FMINOR(device);
199 if (last_maj == maj && last_min == min)
206 * check for this file in our list
208 for_each_file(td, f, i)
209 if (f->major == maj && f->minor == min) {
210 last_fileno = f->fileno;
215 if (lookup_device(td, dev, maj, min)) {
218 dprint(FD_BLKTRACE, "add devices %s\n", dev);
219 fileno = add_file_exclusive(td, dev);
220 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
221 last_fileno = fileno;
228 * Store blk_io_trace data in an ipo for later retrieval.
230 static void store_ipo(struct thread_data *td, unsigned long long offset,
231 unsigned int bytes, int rw, unsigned long long ttime,
234 struct io_piece *ipo = malloc(sizeof(*ipo));
239 * the 512 is wrong here, it should be the hardware sector size...
241 ipo->offset = offset * 512;
243 ipo->delay = ttime / 1000;
245 ipo->ddir = DDIR_WRITE;
247 ipo->ddir = DDIR_READ;
248 ipo->fileno = fileno;
250 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
251 ipo->ddir, ipo->offset,
252 ipo->len, ipo->delay);
253 queue_io_piece(td, ipo);
256 static void handle_trace_notify(struct blk_io_trace *t)
260 log_info("blktrace: got process notify: %x, %d\n",
263 case BLK_TN_TIMESTAMP:
264 log_info("blktrace: got timestamp notify: %x, %d\n",
270 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
275 static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t,
276 unsigned long long ttime, unsigned long *ios)
278 struct io_piece *ipo = malloc(sizeof(*ipo));
282 fileno = trace_add_file(td, t->device);
285 td->o.size += t->bytes;
287 memset(ipo, 0, sizeof(*ipo));
288 INIT_FLIST_HEAD(&ipo->list);
291 * the 512 is wrong here, it should be the hardware sector size...
293 ipo->offset = t->sector * 512;
295 ipo->delay = ttime / 1000;
296 ipo->ddir = DDIR_TRIM;
297 ipo->fileno = fileno;
299 dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
300 ipo->offset, ipo->len,
302 queue_io_piece(td, ipo);
305 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
306 unsigned long long ttime, unsigned long *ios,
312 fileno = trace_add_file(td, t->device);
314 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
316 if (t->bytes > bs[rw])
320 td->o.size += t->bytes;
321 store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
325 * We only care for queue traces, most of the others are side effects
326 * due to internal workings of the block layer.
328 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
329 unsigned long long ttime, unsigned long *ios,
332 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
334 if (t->action & BLK_TC_ACT(BLK_TC_PC))
337 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
338 handle_trace_notify(t);
339 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
340 handle_trace_discard(td, t, ttime, ios);
342 handle_trace_fs(td, t, ttime, ios, bs);
345 static void byteswap_trace(struct blk_io_trace *t)
347 t->magic = fio_swap32(t->magic);
348 t->sequence = fio_swap32(t->sequence);
349 t->time = fio_swap64(t->time);
350 t->sector = fio_swap64(t->sector);
351 t->bytes = fio_swap32(t->bytes);
352 t->action = fio_swap32(t->action);
353 t->pid = fio_swap32(t->pid);
354 t->device = fio_swap32(t->device);
355 t->cpu = fio_swap32(t->cpu);
356 t->error = fio_swap16(t->error);
357 t->pdu_len = fio_swap16(t->pdu_len);
361 * Load a blktrace file by reading all the blk_io_trace entries, and storing
362 * them as io_pieces like the fio text version would do.
364 int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
366 unsigned long long ttime, delay;
367 struct blk_io_trace t;
368 unsigned long ios[2], skipped_writes;
370 unsigned int rw_bs[2];
375 fd = open(filename, O_RDONLY);
377 td_verror(td, errno, "open blktrace file");
381 fifo = fifo_alloc(TRACE_FIFO_SIZE);
388 rw_bs[0] = rw_bs[1] = 0;
391 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
397 else if (ret < (int) sizeof(t)) {
398 log_err("fio: short fifo get\n");
405 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
406 log_err("fio: bad magic in blktrace data: %x\n",
410 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
411 log_err("fio: bad blktrace version %d\n",
415 ret = discard_pdu(td, fifo, fd, &t);
417 td_verror(td, ret, "blktrace lseek");
419 } else if (t.pdu_len != ret) {
420 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
423 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
431 delay = t.time - ttime;
432 if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
436 * set delay to zero if no_stall enabled for
442 handle_trace(td, &t, delay, ios, rw_bs);
449 handle_trace(td, &t, delay, ios, rw_bs);
453 for (i = 0; i < td->files_index; i++) {
455 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
461 if (!td->files_index) {
462 log_err("fio: did not find replay device(s)\n");
467 log_err("fio: %s skips replay of %lu writes due to read-only\n",
468 td->o.name, skipped_writes);
470 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
471 log_err("fio: found no ios in blktrace data\n");
473 } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
474 td->o.td_ddir = TD_DDIR_READ;
475 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
476 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
477 td->o.td_ddir = TD_DDIR_WRITE;
478 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
480 td->o.td_ddir = TD_DDIR_RW;
481 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
482 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
486 * We need to do direct/raw ios to the device, to avoid getting
487 * read-ahead in our way.