* Check if this is a blktrace binary data file. We read a single trace
* into memory and check for the magic signature.
*/
-int is_blktrace(const char *filename)
+int is_blktrace(const char *filename, int *need_swap)
{
struct blk_io_trace t;
int fd, ret;
return 0;
}
- if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
+ if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
+ *need_swap = 0;
return 1;
+ }
+
+ /*
+ * Maybe it needs to be endian swapped...
+ */
+ t.magic = fio_swap32(t.magic);
+ if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
+ *need_swap = 1;
+ return 1;
+ }
return 0;
}
if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
continue;
- sprintf(full_path, "%s/%s", path, dir->d_name);
+ sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name);
if (lstat(full_path, &st) == -1) {
perror("lstat");
break;
*/
if (td->o.replay_redirect) {
dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
- " with: %s", maj, min,
+ " with: %s\n", maj, min,
td->o.replay_redirect);
strcpy(path, td->o.replay_redirect);
found = 1;
#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
-static void trace_add_open_event(struct thread_data *td, int fileno)
+static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
{
struct io_piece *ipo;
ipo = calloc(1, sizeof(*ipo));
+ init_ipo(ipo);
ipo->ddir = DDIR_INVAL;
ipo->fileno = fileno;
- ipo->file_action = FIO_LOG_OPEN_FILE;
+ ipo->file_action = action;
flist_add_tail(&ipo->list, &td->io_log_list);
}
-static void trace_add_file(struct thread_data *td, __u32 device)
+static int trace_add_file(struct thread_data *td, __u32 device)
{
- static unsigned int last_maj, last_min;
+ static unsigned int last_maj, last_min, last_fileno;
unsigned int maj = FMAJOR(device);
unsigned int min = FMINOR(device);
struct fio_file *f;
unsigned int i;
if (last_maj == maj && last_min == min)
- return;
+ return last_fileno;
last_maj = maj;
last_min = min;
* check for this file in our list
*/
for_each_file(td, f, i)
- if (f->major == maj && f->minor == min)
- return;
+ if (f->major == maj && f->minor == min) {
+ last_fileno = f->fileno;
+ return last_fileno;
+ }
strcpy(dev, "/dev");
if (lookup_device(td, dev, maj, min)) {
dprint(FD_BLKTRACE, "add devices %s\n", dev);
fileno = add_file_exclusive(td, dev);
- trace_add_open_event(td, fileno);
+ td->o.open_files++;
+ td->files[fileno]->major = maj;
+ td->files[fileno]->minor = min;
+ trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
+ last_fileno = fileno;
}
+
+ return last_fileno;
}
/*
* Store blk_io_trace data in an ipo for later retrieval.
*/
static void store_ipo(struct thread_data *td, unsigned long long offset,
- unsigned int bytes, int rw, unsigned long long ttime)
+ unsigned int bytes, int rw, unsigned long long ttime,
+ int fileno)
{
struct io_piece *ipo = malloc(sizeof(*ipo));
- memset(ipo, 0, sizeof(*ipo));
- INIT_FLIST_HEAD(&ipo->list);
+ init_ipo(ipo);
+
/*
* the 512 is wrong here, it should be the hardware sector size...
*/
ipo->ddir = DDIR_WRITE;
else
ipo->ddir = DDIR_READ;
+ ipo->fileno = fileno;
dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
ipo->ddir, ipo->offset,
{
switch (t->action) {
case BLK_TN_PROCESS:
- printf("got process notify: %x, %d\n", t->action, t->pid);
+ log_info("blktrace: got process notify: %x, %d\n",
+ t->action, t->pid);
break;
case BLK_TN_TIMESTAMP:
- printf("got timestamp notify: %x, %d\n", t->action, t->pid);
+ log_info("blktrace: got timestamp notify: %x, %d\n",
+ t->action, t->pid);
break;
case BLK_TN_MESSAGE:
break;
unsigned long long ttime, unsigned long *ios)
{
struct io_piece *ipo = malloc(sizeof(*ipo));
+ int fileno;
- trace_add_file(td, t->device);
+ init_ipo(ipo);
+ fileno = trace_add_file(td, t->device);
ios[DDIR_WRITE]++;
td->o.size += t->bytes;
ipo->len = t->bytes;
ipo->delay = ttime / 1000;
ipo->ddir = DDIR_TRIM;
+ ipo->fileno = fileno;
dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
ipo->offset, ipo->len,
unsigned int *bs)
{
int rw;
+ int fileno;
- trace_add_file(td, t->device);
+ fileno = trace_add_file(td, t->device);
rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
ios[rw]++;
td->o.size += t->bytes;
- store_ipo(td, t->sector, t->bytes, rw, ttime);
+ store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
}
/*
handle_trace_fs(td, t, ttime, ios, bs);
}
+static void byteswap_trace(struct blk_io_trace *t)
+{
+ t->magic = fio_swap32(t->magic);
+ t->sequence = fio_swap32(t->sequence);
+ t->time = fio_swap64(t->time);
+ t->sector = fio_swap64(t->sector);
+ t->bytes = fio_swap32(t->bytes);
+ t->action = fio_swap32(t->action);
+ t->pid = fio_swap32(t->pid);
+ t->device = fio_swap32(t->device);
+ t->cpu = fio_swap32(t->cpu);
+ t->error = fio_swap16(t->error);
+ t->pdu_len = fio_swap16(t->pdu_len);
+}
+
/*
* Load a blktrace file by reading all the blk_io_trace entries, and storing
* them as io_pieces like the fio text version would do.
*/
-int load_blktrace(struct thread_data *td, const char *filename)
+int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
{
unsigned long long ttime, delay;
struct blk_io_trace t;
unsigned int cpu;
unsigned int rw_bs[2];
struct fifo *fifo;
- int fd;
+ int fd, i, old_state;
+ struct fio_file *f;
+ int this_depth, depth;
fd = open(filename, O_RDONLY);
if (fd < 0) {
fifo = fifo_alloc(TRACE_FIFO_SIZE);
+ old_state = td_bump_runstate(td, TD_SETTING_UP);
+
td->o.size = 0;
cpu = 0;
ios[0] = ios[1] = 0;
rw_bs[0] = rw_bs[1] = 0;
skipped_writes = 0;
+ this_depth = depth = 0;
do {
int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
break;
}
+ if (need_swap)
+ byteswap_trace(&t);
+
if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
log_err("fio: bad magic in blktrace data: %x\n",
t.magic);
goto err;
}
if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
+ if ((t.action & 0xffff) == __BLK_TA_QUEUE)
+ this_depth++;
+ else if ((t.action & 0xffff) == __BLK_TA_COMPLETE) {
+ depth = max(depth, this_depth);
+ this_depth = 0;
+ }
if (!ttime) {
ttime = t.time;
cpu = t.cpu;
}
} while (1);
+ for (i = 0; i < td->files_index; i++) {
+ f = td->files[i];
+ trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
+ }
+
fifo_free(fifo);
close(fd);
+ td_restore_runstate(td, old_state);
+
+ if (!td->files_index) {
+ log_err("fio: did not find replay device(s)\n");
+ return 1;
+ }
+
+ /*
+ * For stacked devices, we don't always get a COMPLETE event so
+ * the depth grows to insane values. Limit it to something sane(r).
+ */
+ if (!depth || depth > 1024)
+ depth = 1024;
+
if (skipped_writes)
log_err("fio: %s skips replay of %lu writes due to read-only\n",
td->o.name, skipped_writes);
*/
td->o.odirect = 1;
+ /*
+ * we don't know if this option was set or not. it defaults to 1,
+ * so we'll just guess that we should override it if it's still 1
+ */
+ if (td->o.iodepth != 1)
+ td->o.iodepth = depth;
+
return 0;
err:
close(fd);