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(char *path, unsigned int maj, unsigned int min)
111 while ((dir = readdir(D)) != NULL) {
114 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
117 sprintf(full_path, "%s/%s", path, dir->d_name);
118 if (lstat(full_path, &st) == -1) {
123 if (S_ISDIR(st.st_mode)) {
124 found = lookup_device(full_path, maj, min);
126 strcpy(path, full_path);
131 if (!S_ISBLK(st.st_mode))
134 if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
135 dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
136 strcpy(path, full_path);
146 #define FMINORBITS 20
147 #define FMINORMASK ((1U << FMINORBITS) - 1)
148 #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
149 #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
151 static void trace_add_open_event(struct thread_data *td, int fileno)
153 struct io_piece *ipo;
155 ipo = calloc(1, sizeof(*ipo));
157 ipo->ddir = DDIR_INVAL;
158 ipo->fileno = fileno;
159 ipo->file_action = FIO_LOG_OPEN_FILE;
160 flist_add_tail(&ipo->list, &td->io_log_list);
163 static void trace_add_file(struct thread_data *td, __u32 device)
165 static unsigned int last_maj, last_min;
166 unsigned int maj = FMAJOR(device);
167 unsigned int min = FMINOR(device);
172 if (last_maj == maj && last_min == min)
179 * check for this file in our list
181 for_each_file(td, f, i)
182 if (f->major == maj && f->minor == min)
186 if (lookup_device(dev, maj, min)) {
189 dprint(FD_BLKTRACE, "add devices %s\n", dev);
190 fileno = add_file_exclusive(td, dev);
191 trace_add_open_event(td, fileno);
196 * Store blk_io_trace data in an ipo for later retrieval.
198 static void store_ipo(struct thread_data *td, unsigned long long offset,
199 unsigned int bytes, int rw, unsigned long long ttime)
201 struct io_piece *ipo = malloc(sizeof(*ipo));
203 memset(ipo, 0, sizeof(*ipo));
204 INIT_FLIST_HEAD(&ipo->list);
206 * the 512 is wrong here, it should be the hardware sector size...
208 ipo->offset = offset * 512;
210 ipo->delay = ttime / 1000;
212 ipo->ddir = DDIR_WRITE;
214 ipo->ddir = DDIR_READ;
216 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
217 ipo->ddir, ipo->offset,
218 ipo->len, ipo->delay);
219 queue_io_piece(td, ipo);
222 static void handle_trace_notify(struct blk_io_trace *t)
226 printf("got process notify: %x, %d\n", t->action, t->pid);
228 case BLK_TN_TIMESTAMP:
229 printf("got timestamp notify: %x, %d\n", t->action, t->pid);
234 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
239 static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t,
240 unsigned long long ttime, unsigned long *ios)
242 struct io_piece *ipo = malloc(sizeof(*ipo));
244 trace_add_file(td, t->device);
247 td->o.size += t->bytes;
249 memset(ipo, 0, sizeof(*ipo));
250 INIT_FLIST_HEAD(&ipo->list);
253 * the 512 is wrong here, it should be the hardware sector size...
255 ipo->offset = t->sector * 512;
257 ipo->delay = ttime / 1000;
258 ipo->ddir = DDIR_TRIM;
260 dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
261 ipo->offset, ipo->len,
263 queue_io_piece(td, ipo);
266 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
267 unsigned long long ttime, unsigned long *ios,
272 trace_add_file(td, t->device);
274 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
276 if (t->bytes > bs[rw])
280 td->o.size += t->bytes;
281 store_ipo(td, t->sector, t->bytes, rw, ttime);
285 * We only care for queue traces, most of the others are side effects
286 * due to internal workings of the block layer.
288 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
289 unsigned long long ttime, unsigned long *ios,
292 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
294 if (t->action & BLK_TC_ACT(BLK_TC_PC))
297 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
298 handle_trace_notify(t);
299 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
300 handle_trace_discard(td, t, ttime, ios);
302 handle_trace_fs(td, t, ttime, ios, bs);
306 * Load a blktrace file by reading all the blk_io_trace entries, and storing
307 * them as io_pieces like the fio text version would do.
309 int load_blktrace(struct thread_data *td, const char *filename)
311 unsigned long long ttime, delay;
312 struct blk_io_trace t;
313 unsigned long ios[2], skipped_writes;
315 unsigned int rw_bs[2];
319 fd = open(filename, O_RDONLY);
321 td_verror(td, errno, "open blktrace file");
325 fifo = fifo_alloc(TRACE_FIFO_SIZE);
332 rw_bs[0] = rw_bs[1] = 0;
335 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
341 else if (ret < (int) sizeof(t)) {
342 log_err("fio: short fifo get\n");
346 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
347 log_err("fio: bad magic in blktrace data: %x\n",
351 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
352 log_err("fio: bad blktrace version %d\n",
356 ret = discard_pdu(td, fifo, fd, &t);
358 td_verror(td, ret, "blktrace lseek");
360 } else if (t.pdu_len != ret) {
361 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
364 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
372 delay = t.time - ttime;
373 if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
377 * set delay to zero if no_stall enabled for
383 handle_trace(td, &t, delay, ios, rw_bs);
390 handle_trace(td, &t, delay, ios, rw_bs);
398 log_err("fio: %s skips replay of %lu writes due to read-only\n",
399 td->o.name, skipped_writes);
401 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
402 log_err("fio: found no ios in blktrace data\n");
404 } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
405 td->o.td_ddir = TD_DDIR_READ;
406 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
407 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
408 td->o.td_ddir = TD_DDIR_WRITE;
409 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
411 td->o.td_ddir = TD_DDIR_RW;
412 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
413 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
417 * We need to do direct/raw ios to the device, to avoid getting
418 * read-ahead in our way.