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", path, 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));
172 ipo->ddir = DDIR_INVAL;
173 ipo->fileno = fileno;
174 ipo->file_action = FIO_LOG_OPEN_FILE;
175 flist_add_tail(&ipo->list, &td->io_log_list);
178 static void trace_add_file(struct thread_data *td, __u32 device)
180 static unsigned int last_maj, last_min;
181 unsigned int maj = FMAJOR(device);
182 unsigned int min = FMINOR(device);
187 if (last_maj == maj && last_min == min)
194 * check for this file in our list
196 for_each_file(td, f, i)
197 if (f->major == maj && f->minor == min)
201 if (lookup_device(td, dev, maj, min)) {
204 dprint(FD_BLKTRACE, "add devices %s\n", dev);
205 fileno = add_file_exclusive(td, dev);
206 trace_add_open_event(td, fileno);
211 * Store blk_io_trace data in an ipo for later retrieval.
213 static void store_ipo(struct thread_data *td, unsigned long long offset,
214 unsigned int bytes, int rw, unsigned long long ttime)
216 struct io_piece *ipo = malloc(sizeof(*ipo));
218 memset(ipo, 0, sizeof(*ipo));
219 INIT_FLIST_HEAD(&ipo->list);
221 * the 512 is wrong here, it should be the hardware sector size...
223 ipo->offset = offset * 512;
225 ipo->delay = ttime / 1000;
227 ipo->ddir = DDIR_WRITE;
229 ipo->ddir = DDIR_READ;
231 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
232 ipo->ddir, ipo->offset,
233 ipo->len, ipo->delay);
234 queue_io_piece(td, ipo);
237 static void handle_trace_notify(struct blk_io_trace *t)
241 printf("got process notify: %x, %d\n", t->action, t->pid);
243 case BLK_TN_TIMESTAMP:
244 printf("got timestamp notify: %x, %d\n", t->action, t->pid);
249 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
254 static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t,
255 unsigned long long ttime, unsigned long *ios)
257 struct io_piece *ipo = malloc(sizeof(*ipo));
259 trace_add_file(td, t->device);
262 td->o.size += t->bytes;
264 memset(ipo, 0, sizeof(*ipo));
265 INIT_FLIST_HEAD(&ipo->list);
268 * the 512 is wrong here, it should be the hardware sector size...
270 ipo->offset = t->sector * 512;
272 ipo->delay = ttime / 1000;
273 ipo->ddir = DDIR_TRIM;
275 dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
276 ipo->offset, ipo->len,
278 queue_io_piece(td, ipo);
281 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
282 unsigned long long ttime, unsigned long *ios,
287 trace_add_file(td, t->device);
289 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
291 if (t->bytes > bs[rw])
295 td->o.size += t->bytes;
296 store_ipo(td, t->sector, t->bytes, rw, ttime);
300 * We only care for queue traces, most of the others are side effects
301 * due to internal workings of the block layer.
303 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
304 unsigned long long ttime, unsigned long *ios,
307 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
309 if (t->action & BLK_TC_ACT(BLK_TC_PC))
312 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
313 handle_trace_notify(t);
314 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
315 handle_trace_discard(td, t, ttime, ios);
317 handle_trace_fs(td, t, ttime, ios, bs);
321 * Load a blktrace file by reading all the blk_io_trace entries, and storing
322 * them as io_pieces like the fio text version would do.
324 int load_blktrace(struct thread_data *td, const char *filename)
326 unsigned long long ttime, delay;
327 struct blk_io_trace t;
328 unsigned long ios[2], skipped_writes;
330 unsigned int rw_bs[2];
334 fd = open(filename, O_RDONLY);
336 td_verror(td, errno, "open blktrace file");
340 fifo = fifo_alloc(TRACE_FIFO_SIZE);
347 rw_bs[0] = rw_bs[1] = 0;
350 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
356 else if (ret < (int) sizeof(t)) {
357 log_err("fio: short fifo get\n");
361 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
362 log_err("fio: bad magic in blktrace data: %x\n",
366 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
367 log_err("fio: bad blktrace version %d\n",
371 ret = discard_pdu(td, fifo, fd, &t);
373 td_verror(td, ret, "blktrace lseek");
375 } else if (t.pdu_len != ret) {
376 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
379 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
387 delay = t.time - ttime;
388 if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
392 * set delay to zero if no_stall enabled for
398 handle_trace(td, &t, delay, ios, rw_bs);
405 handle_trace(td, &t, delay, ios, rw_bs);
413 log_err("fio: %s skips replay of %lu writes due to read-only\n",
414 td->o.name, skipped_writes);
416 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
417 log_err("fio: found no ios in blktrace data\n");
419 } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
420 td->o.td_ddir = TD_DDIR_READ;
421 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
422 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
423 td->o.td_ddir = TD_DDIR_WRITE;
424 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
426 td->o.td_ddir = TD_DDIR_RW;
427 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
428 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
432 * We need to do direct/raw ios to the device, to avoid getting
433 * read-ahead in our way.