[fio.git] / blktrace.c

/*
 * blktrace support code for fio
 */
#include <stdio.h>
#include <stdlib.h>
#include <sys/stat.h>
#include <dirent.h>

#include "flist.h"
#include "fio.h"
#include "blktrace_api.h"

#define TRACE_FIFO_SIZE	8192

/*
 * fifo refill frontend, to avoid reading data in trace sized bites
 */
static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
{
	char buf[TRACE_FIFO_SIZE];
	unsigned int total;
	int ret;

	total = sizeof(buf);
	if (total > fifo_room(fifo))
		total = fifo_room(fifo);

	ret = read(fd, buf, total);
	if (ret < 0) {
		td_verror(td, errno, "read blktrace file");
		return -1;
	}

	if (ret > 0)
		ret = fifo_put(fifo, buf, ret);

	dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
	return ret;
}

/*
 * Retrieve 'len' bytes from the fifo, refilling if necessary.
 */
static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
			  void *buf, unsigned int len)
{
	if (fifo_len(fifo) < len) {
		int ret = refill_fifo(td, fifo, fd);

		if (ret < 0)
			return ret;
	}

	return fifo_get(fifo, buf, len);
}

/*
 * Just discard the pdu by seeking past it.
 */
static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
		       struct blk_io_trace *t)
{
	if (t->pdu_len == 0)
		return 0;

	dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
	return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
}

/*
 * 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 *need_swap)
{
	struct blk_io_trace t;
	int fd, ret;

	fd = open(filename, O_RDONLY);
	if (fd < 0)
		return 0;

	ret = read(fd, &t, sizeof(t));
	close(fd);

	if (ret < 0) {
		perror("read blktrace");
		return 0;
	} else if (ret != sizeof(t)) {
		log_err("fio: short read on blktrace file\n");
		return 0;
	}

	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;
}

static int lookup_device(struct thread_data *td, char *path, unsigned int maj,
			 unsigned int min)
{
	struct dirent *dir;
	struct stat st;
	int found = 0;
	DIR *D;

	D = opendir(path);
	if (!D)
		return 0;

	while ((dir = readdir(D)) != NULL) {
		char full_path[256];

		if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
			continue;

		sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name);
		if (lstat(full_path, &st) == -1) {
			perror("lstat");
			break;
		}

		if (S_ISDIR(st.st_mode)) {
			found = lookup_device(td, full_path, maj, min);
			if (found) {
				strcpy(path, full_path);
				break;
			}
		}

		if (!S_ISBLK(st.st_mode))
			continue;

		/*
		 * If replay_redirect is set then always return this device
		 * upon lookup which overrides the device lookup based on
		 * major minor in the actual blktrace
		 */
		if (td->o.replay_redirect) {
			dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
					" with: %s\n", maj, min,
					td->o.replay_redirect);
			strcpy(path, td->o.replay_redirect);
			found = 1;
			break;
		}

		if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
			dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
			strcpy(path, full_path);
			found = 1;
			break;
		}
	}

	closedir(D);
	return found;
}

#define FMINORBITS	20
#define FMINORMASK	((1U << FMINORBITS) - 1)
#define FMAJOR(dev)	((unsigned int) ((dev) >> FMINORBITS))
#define FMINOR(dev)	((unsigned int) ((dev) & FMINORMASK))

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 = action;
	flist_add_tail(&ipo->list, &td->io_log_list);
}

static int trace_add_file(struct thread_data *td, __u32 device)
{
	static unsigned int last_maj, last_min, last_fileno;
	unsigned int maj = FMAJOR(device);
	unsigned int min = FMINOR(device);
	struct fio_file *f;
	char dev[256];
	unsigned int i;

	if (last_maj == maj && last_min == min)
		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) {
			last_fileno = f->fileno;
			return last_fileno;
		}

	strcpy(dev, "/dev");
	if (lookup_device(td, dev, maj, min)) {
		int fileno;

		dprint(FD_BLKTRACE, "add devices %s\n", dev);
		fileno = add_file_exclusive(td, dev);
		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,
		      int fileno)
{
	struct io_piece *ipo = malloc(sizeof(*ipo));

	init_ipo(ipo);

	/*
	 * the 512 is wrong here, it should be the hardware sector size...
	 */
	ipo->offset = offset * 512;
	ipo->len = bytes;
	ipo->delay = ttime / 1000;
	if (rw)
		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,
							ipo->len, ipo->delay);
	queue_io_piece(td, ipo);
}

static void handle_trace_notify(struct blk_io_trace *t)
{
	switch (t->action) {
	case BLK_TN_PROCESS:
		log_info("blktrace: got process notify: %x, %d\n",
				t->action, t->pid);
		break;
	case BLK_TN_TIMESTAMP:
		log_info("blktrace: got timestamp notify: %x, %d\n",
				t->action, t->pid);
		break;
	case BLK_TN_MESSAGE:
		break;
	default:
		dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
		break;
	}
}

static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t,
				 unsigned long long ttime, unsigned long *ios)
{
	struct io_piece *ipo = malloc(sizeof(*ipo));
	int fileno;

	init_ipo(ipo);
	fileno = trace_add_file(td, t->device);

	ios[DDIR_WRITE]++;
	td->o.size += t->bytes;

	memset(ipo, 0, sizeof(*ipo));
	INIT_FLIST_HEAD(&ipo->list);

	/*
	 * the 512 is wrong here, it should be the hardware sector size...
	 */
	ipo->offset = t->sector * 512;
	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,
							ipo->delay);
	queue_io_piece(td, ipo);
}

static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
			    unsigned long long ttime, unsigned long *ios,
			    unsigned int *bs)
{
	int rw;
	int fileno;

	fileno = trace_add_file(td, t->device);

	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;

	if (t->bytes > bs[rw])
		bs[rw] = t->bytes;

	ios[rw]++;
	td->o.size += t->bytes;
	store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
}

/*
 * We only care for queue traces, most of the others are side effects
 * due to internal workings of the block layer.
 */
static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
			 unsigned long long ttime, unsigned long *ios,
			 unsigned int *bs)
{
	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
		return;
	if (t->action & BLK_TC_ACT(BLK_TC_PC))
		return;

	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
		handle_trace_notify(t);
	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
		handle_trace_discard(td, t, ttime, ios);
	else
		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 need_swap)
{
	unsigned long long ttime, delay;
	struct blk_io_trace t;
	unsigned long ios[2], skipped_writes;
	unsigned int cpu;
	unsigned int rw_bs[2];
	struct fifo *fifo;
	int fd, i, old_state;
	struct fio_file *f;
	int this_depth, depth;

	fd = open(filename, O_RDONLY);
	if (fd < 0) {
		td_verror(td, errno, "open blktrace file");
		return 1;
	}

	fifo = fifo_alloc(TRACE_FIFO_SIZE);

	old_state = td_bump_runstate(td, TD_SETTING_UP);

	td->o.size = 0;

	cpu = 0;
	ttime = 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));

		if (ret < 0)
			goto err;
		else if (!ret)
			break;
		else if (ret < (int) sizeof(t)) {
			log_err("fio: short fifo get\n");
			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.magic & 0xff) != BLK_IO_TRACE_VERSION) {
			log_err("fio: bad blktrace version %d\n",
								t.magic & 0xff);
			goto err;
		}
		ret = discard_pdu(td, fifo, fd, &t);
		if (ret < 0) {
			td_verror(td, ret, "blktrace lseek");
			goto err;
		} else if (t.pdu_len != ret) {
			log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
			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;
			}

			delay = 0;
			if (cpu == t.cpu)
				delay = t.time - ttime;
			if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
				skipped_writes++;
			else {
				/*
				 * set delay to zero if no_stall enabled for
				 * fast replay
				 */
				if (td->o.no_stall)
					delay = 0;

				handle_trace(td, &t, delay, ios, rw_bs);
			}

			ttime = t.time;
			cpu = t.cpu;
		} else {
			delay = 0;
			handle_trace(td, &t, delay, ios, rw_bs);
		}
	} 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);

	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
		log_err("fio: found no ios in blktrace data\n");
		return 1;
	} else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
		td->o.td_ddir = TD_DDIR_READ;
		td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
		td->o.td_ddir = TD_DDIR_WRITE;
		td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	} else {
		td->o.td_ddir = TD_DDIR_RW;
		td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
		td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	}

	/*
	 * We need to do direct/raw ios to the device, to avoid getting
	 * read-ahead in our way.
	 */
	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);
	fifo_free(fifo);
	return 1;
}
Commit	Line	Data
fb7b71a3 JA	1	/*
	2	* blktrace support code for fio
	3	*/
	4	#include <stdio.h>
	5	#include <stdlib.h>
5e6c2067 JA	6	#include <sys/stat.h>
5e6c2067 JA	7	#include <dirent.h>
8c1fdf04	8
01743ee1	9	#include "flist.h"
fb7b71a3 JA	10	#include "fio.h"
	11	#include "blktrace_api.h"
	12
2da7df1d	13	#define TRACE_FIFO_SIZE 8192
e2887563 JA	14
	15	/*
	16	* fifo refill frontend, to avoid reading data in trace sized bites
	17	*/
	18	static int refill_fifo(struct thread_data td, struct fifo fifo, int fd)
	19	{
	20	char buf[TRACE_FIFO_SIZE];
f12b323f	21	unsigned int total;
e2887563 JA	22	int ret;
e2887563 JA	23
f12b323f JA	24	total = sizeof(buf);
	25	if (total > fifo_room(fifo))
	26	total = fifo_room(fifo);
e2887563	27
f12b323f JA	28	ret = read(fd, buf, total);
	29	if (ret < 0) {
	30	td_verror(td, errno, "read blktrace file");
	31	return -1;
e2887563 JA	32	}
e2887563 JA	33
f12b323f JA	34	if (ret > 0)
	35	ret = fifo_put(fifo, buf, ret);
	36
bd6f78b2	37	dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
f12b323f	38	return ret;
e2887563 JA	39	}
	40
	41	/*
	42	* Retrieve 'len' bytes from the fifo, refilling if necessary.
	43	*/
	44	static int trace_fifo_get(struct thread_data td, struct fifo fifo, int fd,
	45	void *buf, unsigned int len)
	46	{
f12b323f JA	47	if (fifo_len(fifo) < len) {
f12b323f JA	48	int ret = refill_fifo(td, fifo, fd);
e2887563	49
f12b323f JA	50	if (ret < 0)
	51	return ret;
	52	}
e2887563 JA	53
	54	return fifo_get(fifo, buf, len);
	55	}
	56
8c1fdf04 JA	57	/*
	58	* Just discard the pdu by seeking past it.
	59	*/
f12b323f JA	60	static int discard_pdu(struct thread_data td, struct fifo fifo, int fd,
f12b323f JA	61	struct blk_io_trace *t)
fb7b71a3 JA	62	{
	63	if (t->pdu_len == 0)
	64	return 0;
	65
bd6f78b2	66	dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
f12b323f	67	return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
fb7b71a3 JA	68	}
fb7b71a3 JA	69
8c1fdf04 JA	70	/*
	71	* Check if this is a blktrace binary data file. We read a single trace
	72	* into memory and check for the magic signature.
	73	*/
d95b34a6	74	int is_blktrace(const char filename, int need_swap)
fb7b71a3 JA	75	{
	76	struct blk_io_trace t;
	77	int fd, ret;
	78
	79	fd = open(filename, O_RDONLY);
4dced407	80	if (fd < 0)
fb7b71a3	81	return 0;
fb7b71a3 JA	82
	83	ret = read(fd, &t, sizeof(t));
	84	close(fd);
	85
	86	if (ret < 0) {
	87	perror("read blktrace");
	88	return 0;
	89	} else if (ret != sizeof(t)) {
	90	log_err("fio: short read on blktrace file\n");
	91	return 0;
	92	}
	93
d95b34a6 JA	94	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
d95b34a6 JA	95	*need_swap = 0;
fb7b71a3	96	return 1;
d95b34a6 JA	97	}
	98
	99	/*
	100	* Maybe it needs to be endian swapped...
	101	*/
	102	t.magic = fio_swap32(t.magic);
	103	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
	104	*need_swap = 1;
	105	return 1;
	106	}
fb7b71a3 JA	107
	108	return 0;
	109	}
	110
d1c46c04 DN	111	static int lookup_device(struct thread_data td, char path, unsigned int maj,
d1c46c04 DN	112	unsigned int min)
5e6c2067 JA	113	{
	114	struct dirent *dir;
	115	struct stat st;
	116	int found = 0;
	117	DIR *D;
	118
	119	D = opendir(path);
	120	if (!D)
	121	return 0;
	122
	123	while ((dir = readdir(D)) != NULL) {
	124	char full_path[256];
	125
	126	if (!strcmp(dir->d_name, ".") \|\| !strcmp(dir->d_name, ".."))
	127	continue;
	128
b9fd788f	129	sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name);
5e6c2067 JA	130	if (lstat(full_path, &st) == -1) {
	131	perror("lstat");
	132	break;
	133	}
	134
	135	if (S_ISDIR(st.st_mode)) {
d1c46c04	136	found = lookup_device(td, full_path, maj, min);
5e6c2067 JA	137	if (found) {
	138	strcpy(path, full_path);
	139	break;
	140	}
	141	}
	142
	143	if (!S_ISBLK(st.st_mode))
	144	continue;
	145
d1c46c04 DN	146	/*
	147	* If replay_redirect is set then always return this device
	148	* upon lookup which overrides the device lookup based on
	149	* major minor in the actual blktrace
	150	*/
	151	if (td->o.replay_redirect) {
	152	dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
25a4e78b	153	" with: %s\n", maj, min,
d1c46c04 DN	154	td->o.replay_redirect);
	155	strcpy(path, td->o.replay_redirect);
	156	found = 1;
	157	break;
	158	}
	159
5e6c2067	160	if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
5ec10eaa	161	dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
5e6c2067 JA	162	strcpy(path, full_path);
	163	found = 1;
	164	break;
	165	}
	166	}
	167
	168	closedir(D);
	169	return found;
	170	}
	171
c69aa91f JA	172	#define FMINORBITS 20
	173	#define FMINORMASK ((1U << FMINORBITS) - 1)
	174	#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
	175	#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
eeb9c2aa	176
89ac1d48	177	static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
691c8fb0 JA	178	{
	179	struct io_piece *ipo;
	180
	181	ipo = calloc(1, sizeof(*ipo));
0d29de83	182	init_ipo(ipo);
691c8fb0 JA	183
	184	ipo->ddir = DDIR_INVAL;
	185	ipo->fileno = fileno;
89ac1d48	186	ipo->file_action = action;
01743ee1	187	flist_add_tail(&ipo->list, &td->io_log_list);
691c8fb0 JA	188	}
691c8fb0 JA	189
89ac1d48	190	static int trace_add_file(struct thread_data *td, __u32 device)
5e6c2067	191	{
89ac1d48	192	static unsigned int last_maj, last_min, last_fileno;
c69aa91f JA	193	unsigned int maj = FMAJOR(device);
c69aa91f JA	194	unsigned int min = FMINOR(device);
5e6c2067 JA	195	struct fio_file *f;
	196	char dev[256];
	197	unsigned int i;
	198
	199	if (last_maj == maj && last_min == min)
89ac1d48	200	return last_fileno;
5e6c2067 JA	201
	202	last_maj = maj;
	203	last_min = min;
	204
	205	/*
	206	* check for this file in our list
	207	*/
	208	for_each_file(td, f, i)
89ac1d48 SL	209	if (f->major == maj && f->minor == min) {
	210	last_fileno = f->fileno;
	211	return last_fileno;
	212	}
5e6c2067 JA	213
5e6c2067 JA	214	strcpy(dev, "/dev");
d1c46c04	215	if (lookup_device(td, dev, maj, min)) {
691c8fb0 JA	216	int fileno;
691c8fb0 JA	217
bd6f78b2	218	dprint(FD_BLKTRACE, "add devices %s\n", dev);
49ffb4a2	219	fileno = add_file_exclusive(td, dev);
b53f2c54	220	td->o.open_files++;
5903e7b7 JA	221	td->files[fileno]->major = maj;
5903e7b7 JA	222	td->files[fileno]->minor = min;
89ac1d48 SL	223	trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
89ac1d48 SL	224	last_fileno = fileno;
bd6f78b2	225	}
f01b34ae	226
89ac1d48	227	return last_fileno;
5e6c2067 JA	228	}
5e6c2067 JA	229
8c1fdf04 JA	230	/*
	231	* Store blk_io_trace data in an ipo for later retrieval.
	232	*/
fdefd987	233	static void store_ipo(struct thread_data *td, unsigned long long offset,
89ac1d48 SL	234	unsigned int bytes, int rw, unsigned long long ttime,
89ac1d48 SL	235	int fileno)
fdefd987 JA	236	{
	237	struct io_piece ipo = malloc(sizeof(ipo));
	238
0d29de83 JA	239	init_ipo(ipo);
0d29de83 JA	240
a2eea81b JA	241	/*
	242	* the 512 is wrong here, it should be the hardware sector size...
	243	*/
	244	ipo->offset = offset * 512;
fdefd987	245	ipo->len = bytes;
8c1fdf04	246	ipo->delay = ttime / 1000;
fdefd987 JA	247	if (rw)
	248	ipo->ddir = DDIR_WRITE;
	249	else
	250	ipo->ddir = DDIR_READ;
89ac1d48	251	ipo->fileno = fileno;
fdefd987	252
bd6f78b2 JA	253	dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
	254	ipo->ddir, ipo->offset,
	255	ipo->len, ipo->delay);
691c8fb0	256	queue_io_piece(td, ipo);
fdefd987 JA	257	}
fdefd987 JA	258
0b9d69ec	259	static void handle_trace_notify(struct blk_io_trace *t)
cd991b9e	260	{
691c8fb0 JA	261	switch (t->action) {
691c8fb0 JA	262	case BLK_TN_PROCESS:
d95b34a6 JA	263	log_info("blktrace: got process notify: %x, %d\n",
d95b34a6 JA	264	t->action, t->pid);
691c8fb0 JA	265	break;
691c8fb0 JA	266	case BLK_TN_TIMESTAMP:
d95b34a6 JA	267	log_info("blktrace: got timestamp notify: %x, %d\n",
d95b34a6 JA	268	t->action, t->pid);
691c8fb0	269	break;
ff58fced JA	270	case BLK_TN_MESSAGE:
ff58fced JA	271	break;
691c8fb0 JA	272	default:
	273	dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
	274	break;
	275	}
	276	}
5b3023b8	277
ff58fced JA	278	static void handle_trace_discard(struct thread_data td, struct blk_io_trace t,
	279	unsigned long long ttime, unsigned long *ios)
	280	{
	281	struct io_piece ipo = malloc(sizeof(ipo));
89ac1d48	282	int fileno;
ff58fced	283
0d29de83	284	init_ipo(ipo);
89ac1d48	285	fileno = trace_add_file(td, t->device);
ff58fced JA	286
	287	ios[DDIR_WRITE]++;
	288	td->o.size += t->bytes;
	289
	290	memset(ipo, 0, sizeof(*ipo));
	291	INIT_FLIST_HEAD(&ipo->list);
	292
	293	/*
	294	* the 512 is wrong here, it should be the hardware sector size...
	295	*/
	296	ipo->offset = t->sector * 512;
	297	ipo->len = t->bytes;
	298	ipo->delay = ttime / 1000;
	299	ipo->ddir = DDIR_TRIM;
89ac1d48	300	ipo->fileno = fileno;
ff58fced JA	301
	302	dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
	303	ipo->offset, ipo->len,
	304	ipo->delay);
	305	queue_io_piece(td, ipo);
	306	}
	307
691c8fb0 JA	308	static void handle_trace_fs(struct thread_data td, struct blk_io_trace t,
	309	unsigned long long ttime, unsigned long *ios,
	310	unsigned int *bs)
	311	{
	312	int rw;
89ac1d48	313	int fileno;
5b3023b8	314
89ac1d48	315	fileno = trace_add_file(td, t->device);
5b3023b8 JA	316
	317	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
	318
	319	if (t->bytes > bs[rw])
	320	bs[rw] = t->bytes;
	321
	322	ios[rw]++;
	323	td->o.size += t->bytes;
89ac1d48	324	store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
cd991b9e JA	325	}
cd991b9e JA	326
691c8fb0 JA	327	/*
	328	* We only care for queue traces, most of the others are side effects
	329	* due to internal workings of the block layer.
	330	*/
	331	static void handle_trace(struct thread_data td, struct blk_io_trace t,
	332	unsigned long long ttime, unsigned long *ios,
	333	unsigned int *bs)
	334	{
	335	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
	336	return;
	337	if (t->action & BLK_TC_ACT(BLK_TC_PC))
	338	return;
	339
	340	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
aec2de20	341	handle_trace_notify(t);
ff58fced JA	342	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
ff58fced JA	343	handle_trace_discard(td, t, ttime, ios);
691c8fb0 JA	344	else
	345	handle_trace_fs(td, t, ttime, ios, bs);
	346	}
	347
d95b34a6 JA	348	static void byteswap_trace(struct blk_io_trace *t)
	349	{
	350	t->magic = fio_swap32(t->magic);
	351	t->sequence = fio_swap32(t->sequence);
	352	t->time = fio_swap64(t->time);
	353	t->sector = fio_swap64(t->sector);
	354	t->bytes = fio_swap32(t->bytes);
	355	t->action = fio_swap32(t->action);
	356	t->pid = fio_swap32(t->pid);
	357	t->device = fio_swap32(t->device);
	358	t->cpu = fio_swap32(t->cpu);
	359	t->error = fio_swap16(t->error);
	360	t->pdu_len = fio_swap16(t->pdu_len);
	361	}
	362
8c1fdf04 JA	363	/*
	364	* Load a blktrace file by reading all the blk_io_trace entries, and storing
	365	* them as io_pieces like the fio text version would do.
	366	*/
d95b34a6	367	int load_blktrace(struct thread_data td, const char filename, int need_swap)
fb7b71a3	368	{
a61eddec	369	unsigned long long ttime, delay;
fb7b71a3	370	struct blk_io_trace t;
4241ea8f	371	unsigned long ios[2], skipped_writes;
a61eddec	372	unsigned int cpu;
d84f8d49	373	unsigned int rw_bs[2];
e2887563	374	struct fifo *fifo;
5903e7b7	375	int fd, i, old_state;
89ac1d48	376	struct fio_file *f;
eb5fdcf1	377	int this_depth, depth;
fb7b71a3 JA	378
	379	fd = open(filename, O_RDONLY);
	380	if (fd < 0) {
	381	td_verror(td, errno, "open blktrace file");
	382	return 1;
	383	}
	384
e2887563 JA	385	fifo = fifo_alloc(TRACE_FIFO_SIZE);
e2887563 JA	386
8edd973d	387	old_state = td_bump_runstate(td, TD_SETTING_UP);
5903e7b7	388
6df8adaa JA	389	td->o.size = 0;
6df8adaa JA	390
a61eddec	391	cpu = 0;
d84f8d49 JA	392	ttime = 0;
	393	ios[0] = ios[1] = 0;
	394	rw_bs[0] = rw_bs[1] = 0;
4241ea8f	395	skipped_writes = 0;
eb5fdcf1	396	this_depth = depth = 0;
fb7b71a3	397	do {
e2887563	398	int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
fb7b71a3	399
e2887563	400	if (ret < 0)
8c1fdf04	401	goto err;
e2887563 JA	402	else if (!ret)
	403	break;
	404	else if (ret < (int) sizeof(t)) {
	405	log_err("fio: short fifo get\n");
fb7b71a3	406	break;
fb7b71a3 JA	407	}
fb7b71a3 JA	408
d95b34a6 JA	409	if (need_swap)
	410	byteswap_trace(&t);
	411
fb7b71a3	412	if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
5ec10eaa JA	413	log_err("fio: bad magic in blktrace data: %x\n",
5ec10eaa JA	414	t.magic);
8c1fdf04	415	goto err;
fb7b71a3 JA	416	}
fb7b71a3 JA	417	if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
5ec10eaa JA	418	log_err("fio: bad blktrace version %d\n",
5ec10eaa JA	419	t.magic & 0xff);
8c1fdf04	420	goto err;
fb7b71a3	421	}
f12b323f JA	422	ret = discard_pdu(td, fifo, fd, &t);
f12b323f JA	423	if (ret < 0) {
fb7b71a3	424	td_verror(td, ret, "blktrace lseek");
8c1fdf04	425	goto err;
f12b323f JA	426	} else if (t.pdu_len != ret) {
	427	log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
	428	goto err;
fb7b71a3	429	}
691c8fb0	430	if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
eb5fdcf1 JA	431	if ((t.action & 0xffff) == __BLK_TA_QUEUE)
	432	this_depth++;
	433	else if ((t.action & 0xffff) == __BLK_TA_COMPLETE) {
	434	depth = max(depth, this_depth);
	435	this_depth = 0;
	436	}
691c8fb0 JA	437	if (!ttime) {
	438	ttime = t.time;
	439	cpu = t.cpu;
	440	}
	441
	442	delay = 0;
	443	if (cpu == t.cpu)
	444	delay = t.time - ttime;
	445	if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
	446	skipped_writes++;
64bbb865 DN	447	else {
	448	/*
	449	* set delay to zero if no_stall enabled for
	450	* fast replay
	451	*/
	452	if (td->o.no_stall)
	453	delay = 0;
	454
691c8fb0	455	handle_trace(td, &t, delay, ios, rw_bs);
64bbb865	456	}
691c8fb0	457
8c1fdf04	458	ttime = t.time;
a61eddec	459	cpu = t.cpu;
a6edd638 JA	460	} else {
a6edd638 JA	461	delay = 0;
4241ea8f	462	handle_trace(td, &t, delay, ios, rw_bs);
a6edd638	463	}
fb7b71a3 JA	464	} while (1);
fb7b71a3 JA	465
85a47ca2	466	for (i = 0; i < td->files_index; i++) {
f01b34ae	467	f = td->files[i];
89ac1d48	468	trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
85a47ca2	469	}
89ac1d48	470
38470f85	471	fifo_free(fifo);
fb7b71a3	472	close(fd);
8c1fdf04	473
8edd973d	474	td_restore_runstate(td, old_state);
5903e7b7	475
f01b34ae JA	476	if (!td->files_index) {
	477	log_err("fio: did not find replay device(s)\n");
	478	return 1;
	479	}
	480
eb5fdcf1 JA	481	/*
	482	* For stacked devices, we don't always get a COMPLETE event so
	483	* the depth grows to insane values. Limit it to something sane(r).
	484	*/
	485	if (!depth \|\| depth > 1024)
	486	depth = 1024;
	487
4241ea8f	488	if (skipped_writes)
5ec10eaa JA	489	log_err("fio: %s skips replay of %lu writes due to read-only\n",
5ec10eaa JA	490	td->o.name, skipped_writes);
4241ea8f	491
8c1fdf04 JA	492	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	493	log_err("fio: found no ios in blktrace data\n");
	494	return 1;
d84f8d49	495	} else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
8c1fdf04	496	td->o.td_ddir = TD_DDIR_READ;
d84f8d49 JA	497	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
d84f8d49 JA	498	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
8c1fdf04	499	td->o.td_ddir = TD_DDIR_WRITE;
d84f8d49 JA	500	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
d84f8d49 JA	501	} else {
8c1fdf04	502	td->o.td_ddir = TD_DDIR_RW;
d84f8d49 JA	503	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	504	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	505	}
8c1fdf04 JA	506
	507	/*
	508	* We need to do direct/raw ios to the device, to avoid getting
	509	* read-ahead in our way.
	510	*/
	511	td->o.odirect = 1;
	512
eb5fdcf1 JA	513	/*
	514	* we don't know if this option was set or not. it defaults to 1,
	515	* so we'll just guess that we should override it if it's still 1
	516	*/
	517	if (td->o.iodepth != 1)
	518	td->o.iodepth = depth;
	519
fb7b71a3	520	return 0;
8c1fdf04 JA	521	err:
8c1fdf04 JA	522	close(fd);
38470f85	523	fifo_free(fifo);
8c1fdf04	524	return 1;
fb7b71a3	525	}