2 * blktrace support code for fio
11 #include "blktrace_api.h"
13 #define TRACE_FIFO_SIZE 65536
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);
41 * Retrieve 'len' bytes from the fifo, refilling if necessary.
43 static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
44 void *buf, unsigned int len)
46 if (fifo_len(fifo) < len) {
47 int ret = refill_fifo(td, fifo, fd);
53 return fifo_get(fifo, buf, len);
57 * Just discard the pdu by seeking past it.
59 static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
60 struct blk_io_trace *t)
65 return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
69 * Check if this is a blktrace binary data file. We read a single trace
70 * into memory and check for the magic signature.
72 int is_blktrace(const char *filename)
74 struct blk_io_trace t;
77 fd = open(filename, O_RDONLY);
79 perror("open blktrace");
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 strcpy(path, full_path);
145 #define FMINORBITS 20
146 #define FMINORMASK ((1U << FMINORBITS) - 1)
147 #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
148 #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
150 static void trace_add_file(struct thread_data *td, __u32 device)
152 static unsigned int last_maj, last_min;
153 unsigned int maj = FMAJOR(device);
154 unsigned int min = FMINOR(device);
159 if (last_maj == maj && last_min == min)
166 * check for this file in our list
168 for_each_file(td, f, i)
169 if (f->major == maj && f->minor == min)
173 if (lookup_device(dev, maj, min))
178 * Store blk_io_trace data in an ipo for later retrieval.
180 static void store_ipo(struct thread_data *td, unsigned long long offset,
181 unsigned int bytes, int rw, unsigned long long ttime)
183 struct io_piece *ipo = malloc(sizeof(*ipo));
185 memset(ipo, 0, sizeof(*ipo));
186 INIT_LIST_HEAD(&ipo->list);
188 * the 512 is wrong here, it should be the hardware sector size...
190 ipo->offset = offset * 512;
192 ipo->delay = ttime / 1000;
194 ipo->ddir = DDIR_WRITE;
196 ipo->ddir = DDIR_READ;
198 list_add_tail(&ipo->list, &td->io_log_list);
202 * We only care for queue traces, most of the others are side effects
203 * due to internal workings of the block layer.
205 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
206 unsigned long long ttime, unsigned long *ios,
211 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
213 if (t->action & BLK_TC_ACT(BLK_TC_PC))
215 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
218 trace_add_file(td, t->device);
220 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
222 if (t->bytes > bs[rw])
226 td->o.size += t->bytes;
227 store_ipo(td, t->sector, t->bytes, rw, ttime);
231 * Load a blktrace file by reading all the blk_io_trace entries, and storing
232 * them as io_pieces like the fio text version would do.
234 int load_blktrace(struct thread_data *td, const char *filename)
236 unsigned long long ttime, delay;
237 struct blk_io_trace t;
238 unsigned long ios[2];
240 unsigned int rw_bs[2];
244 fd = open(filename, O_RDONLY);
246 td_verror(td, errno, "open blktrace file");
250 fifo = fifo_alloc(TRACE_FIFO_SIZE);
257 rw_bs[0] = rw_bs[1] = 0;
259 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
265 else if (ret < (int) sizeof(t)) {
266 log_err("fio: short fifo get\n");
270 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
271 log_err("fio: bad magic in blktrace data: %x\n", t.magic);
274 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
275 log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
278 ret = discard_pdu(td, fifo, fd, &t);
280 td_verror(td, ret, "blktrace lseek");
282 } else if (t.pdu_len != ret) {
283 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
292 delay = t.time - ttime;
293 handle_trace(td, &t, delay, ios, rw_bs);
301 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
302 log_err("fio: found no ios in blktrace data\n");
304 } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
305 td->o.td_ddir = TD_DDIR_READ;
306 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
307 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
308 td->o.td_ddir = TD_DDIR_WRITE;
309 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
311 td->o.td_ddir = TD_DDIR_RW;
312 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
313 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
317 * We need to do direct/raw ios to the device, to avoid getting
318 * read-ahead in our way.