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)
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 static int lookup_device(struct thread_data *td, char *path, unsigned int maj,
112 while ((dir = readdir(D)) != NULL) {
115 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
118 sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name);
119 if (lstat(full_path, &st) == -1) {
124 if (S_ISDIR(st.st_mode)) {
125 found = lookup_device(td, full_path, maj, min);
127 strcpy(path, full_path);
132 if (!S_ISBLK(st.st_mode))
136 * If replay_redirect is set then always return this device
137 * upon lookup which overrides the device lookup based on
138 * major minor in the actual blktrace
140 if (td->o.replay_redirect) {
141 dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
142 " with: %s", maj, min,
143 td->o.replay_redirect);
144 strcpy(path, td->o.replay_redirect);
149 if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
150 dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
151 strcpy(path, full_path);
161 #define FMINORBITS 20
162 #define FMINORMASK ((1U << FMINORBITS) - 1)
163 #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
164 #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
166 static void trace_add_open_event(struct thread_data *td, int fileno)
168 struct io_piece *ipo;
170 ipo = calloc(1, sizeof(*ipo));
173 ipo->ddir = DDIR_INVAL;
174 ipo->fileno = fileno;
175 ipo->file_action = FIO_LOG_OPEN_FILE;
176 flist_add_tail(&ipo->list, &td->io_log_list);
179 static void trace_add_file(struct thread_data *td, __u32 device)
181 static unsigned int last_maj, last_min;
182 unsigned int maj = FMAJOR(device);
183 unsigned int min = FMINOR(device);
188 if (last_maj == maj && last_min == min)
195 * check for this file in our list
197 for_each_file(td, f, i)
198 if (f->major == maj && f->minor == min)
202 if (lookup_device(td, dev, maj, min)) {
205 dprint(FD_BLKTRACE, "add devices %s\n", dev);
206 fileno = add_file_exclusive(td, dev);
207 trace_add_open_event(td, fileno);
212 * Store blk_io_trace data in an ipo for later retrieval.
214 static void store_ipo(struct thread_data *td, unsigned long long offset,
215 unsigned int bytes, int rw, unsigned long long ttime)
217 struct io_piece *ipo = malloc(sizeof(*ipo));
222 * the 512 is wrong here, it should be the hardware sector size...
224 ipo->offset = offset * 512;
226 ipo->delay = ttime / 1000;
228 ipo->ddir = DDIR_WRITE;
230 ipo->ddir = DDIR_READ;
232 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
233 ipo->ddir, ipo->offset,
234 ipo->len, ipo->delay);
235 queue_io_piece(td, ipo);
238 static void handle_trace_notify(struct blk_io_trace *t)
242 printf("got process notify: %x, %d\n", t->action, t->pid);
244 case BLK_TN_TIMESTAMP:
245 printf("got timestamp notify: %x, %d\n", t->action, t->pid);
250 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
255 static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t,
256 unsigned long long ttime, unsigned long *ios)
258 struct io_piece *ipo = malloc(sizeof(*ipo));
261 trace_add_file(td, t->device);
264 td->o.size += t->bytes;
266 memset(ipo, 0, sizeof(*ipo));
267 INIT_FLIST_HEAD(&ipo->list);
270 * the 512 is wrong here, it should be the hardware sector size...
272 ipo->offset = t->sector * 512;
274 ipo->delay = ttime / 1000;
275 ipo->ddir = DDIR_TRIM;
277 dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
278 ipo->offset, ipo->len,
280 queue_io_piece(td, ipo);
283 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
284 unsigned long long ttime, unsigned long *ios,
289 trace_add_file(td, t->device);
291 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
293 if (t->bytes > bs[rw])
297 td->o.size += t->bytes;
298 store_ipo(td, t->sector, t->bytes, rw, ttime);
302 * We only care for queue traces, most of the others are side effects
303 * due to internal workings of the block layer.
305 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
306 unsigned long long ttime, unsigned long *ios,
309 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
311 if (t->action & BLK_TC_ACT(BLK_TC_PC))
314 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
315 handle_trace_notify(t);
316 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
317 handle_trace_discard(td, t, ttime, ios);
319 handle_trace_fs(td, t, ttime, ios, bs);
323 * Load a blktrace file by reading all the blk_io_trace entries, and storing
324 * them as io_pieces like the fio text version would do.
326 int load_blktrace(struct thread_data *td, const char *filename)
328 unsigned long long ttime, delay;
329 struct blk_io_trace t;
330 unsigned long ios[2], skipped_writes;
332 unsigned int rw_bs[2];
336 fd = open(filename, O_RDONLY);
338 td_verror(td, errno, "open blktrace file");
342 fifo = fifo_alloc(TRACE_FIFO_SIZE);
349 rw_bs[0] = rw_bs[1] = 0;
352 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
358 else if (ret < (int) sizeof(t)) {
359 log_err("fio: short fifo get\n");
363 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
364 log_err("fio: bad magic in blktrace data: %x\n",
368 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
369 log_err("fio: bad blktrace version %d\n",
373 ret = discard_pdu(td, fifo, fd, &t);
375 td_verror(td, ret, "blktrace lseek");
377 } else if (t.pdu_len != ret) {
378 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
381 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
389 delay = t.time - ttime;
390 if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
394 * set delay to zero if no_stall enabled for
400 handle_trace(td, &t, delay, ios, rw_bs);
407 handle_trace(td, &t, delay, ios, rw_bs);
415 log_err("fio: %s skips replay of %lu writes due to read-only\n",
416 td->o.name, skipped_writes);
418 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
419 log_err("fio: found no ios in blktrace data\n");
421 } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
422 td->o.td_ddir = TD_DDIR_READ;
423 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
424 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
425 td->o.td_ddir = TD_DDIR_WRITE;
426 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
428 td->o.td_ddir = TD_DDIR_RW;
429 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
430 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
434 * We need to do direct/raw ios to the device, to avoid getting
435 * read-ahead in our way.