[fio.git] / blktrace.c

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

#include "list.h"
#include "fio.h"
#include "blktrace_api.h"

#define TRACE_FIFO_SIZE	(sizeof(struct blk_io_trace) * 1000)

/*
 * 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, left;
	void *ptr;
	int ret;

	total = 0;
	ptr = buf;
	while (total < TRACE_FIFO_SIZE) {
		left = TRACE_FIFO_SIZE - total;

		ret = read(fd, ptr, left);
		if (ret < 0) {
			td_verror(td, errno, "read blktrace file");
			return -1;
		} else if (!ret)
			break;

		fifo_put(fifo, ptr, ret);
		ptr += ret;
		total += ret;
	}

	return 0;
}

/*
 * 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)
{
	int ret;

	if (fifo_len(fifo) >= len)
		return fifo_get(fifo, buf, len);

	ret = refill_fifo(td, fifo, fd);
	if (ret < 0)
		return ret;

	if (fifo_len(fifo) < len)
		return 0;

	return fifo_get(fifo, buf, len);
}

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

	if (lseek(fd, t->pdu_len, SEEK_CUR) < 0)
		return errno;
		
	return 0;
}

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

	fd = open(filename, O_RDONLY);
	if (fd < 0) {
		perror("open blktrace");
		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)
		return 1;

	return 0;
}

/*
 * 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)
{
	struct io_piece *ipo = malloc(sizeof(*ipo));

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

	list_add_tail(&ipo->list, &td->io_log_list);
}

/*
 * 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)
{
	int rw;

	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
		return;
	if (t->action & BLK_TC_ACT(BLK_TC_PC))
		return;

	/*
	 * should not happen, need to look into that...
	 */
	if (!t->bytes)
		return;

	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);
}

/*
 * 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)
{
	unsigned long long ttime, delay;
	struct blk_io_trace t;
	unsigned long ios[2];
	unsigned int cpu;
	unsigned int rw_bs[2];
	struct fifo *fifo;
	int fd;

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

	fifo = fifo_alloc(TRACE_FIFO_SIZE);

	td->o.size = 0;

	cpu = 0;
	ttime = 0;
	ios[0] = ios[1] = 0;
	rw_bs[0] = rw_bs[1] = 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 ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
			log_err("fio: bad magic in blktrace data\n");
			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(fd, &t);
		if (ret) {
			td_verror(td, ret, "blktrace lseek");
			goto err;
		}
		if (!ttime) {
			ttime = t.time;
			cpu = t.cpu;
		}
		delay = 0;
		if (cpu == t.cpu)
			delay = t.time - ttime;
		handle_trace(td, &t, delay, ios, rw_bs);
		ttime = t.time;
		cpu = t.cpu;
	} while (1);

	fifo_free(fifo);
	close(fd);

	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;

	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>
8c1fdf04	6
fb7b71a3 JA	7	#include "list.h"
	8	#include "fio.h"
	9	#include "blktrace_api.h"
	10
e2887563 JA	11	#define TRACE_FIFO_SIZE (sizeof(struct blk_io_trace) * 1000)
	12
	13	/*
	14	* fifo refill frontend, to avoid reading data in trace sized bites
	15	*/
	16	static int refill_fifo(struct thread_data td, struct fifo fifo, int fd)
	17	{
	18	char buf[TRACE_FIFO_SIZE];
	19	unsigned int total, left;
	20	void *ptr;
	21	int ret;
	22
	23	total = 0;
	24	ptr = buf;
	25	while (total < TRACE_FIFO_SIZE) {
	26	left = TRACE_FIFO_SIZE - total;
	27
	28	ret = read(fd, ptr, left);
	29	if (ret < 0) {
	30	td_verror(td, errno, "read blktrace file");
	31	return -1;
	32	} else if (!ret)
	33	break;
	34
	35	fifo_put(fifo, ptr, ret);
	36	ptr += ret;
	37	total += ret;
	38	}
	39
	40	return 0;
	41	}
	42
	43	/*
	44	* Retrieve 'len' bytes from the fifo, refilling if necessary.
	45	*/
	46	static int trace_fifo_get(struct thread_data td, struct fifo fifo, int fd,
	47	void *buf, unsigned int len)
	48	{
	49	int ret;
	50
	51	if (fifo_len(fifo) >= len)
	52	return fifo_get(fifo, buf, len);
	53
	54	ret = refill_fifo(td, fifo, fd);
	55	if (ret < 0)
	56	return ret;
	57
	58	if (fifo_len(fifo) < len)
	59	return 0;
	60
	61	return fifo_get(fifo, buf, len);
	62	}
	63
8c1fdf04 JA	64	/*
	65	* Just discard the pdu by seeking past it.
	66	*/
fb7b71a3 JA	67	static int discard_pdu(int fd, struct blk_io_trace *t)
	68	{
	69	if (t->pdu_len == 0)
	70	return 0;
	71
	72	if (lseek(fd, t->pdu_len, SEEK_CUR) < 0)
	73	return errno;
	74
	75	return 0;
	76	}
	77
8c1fdf04 JA	78	/*
	79	* Check if this is a blktrace binary data file. We read a single trace
	80	* into memory and check for the magic signature.
	81	*/
fb7b71a3 JA	82	int is_blktrace(const char *filename)
	83	{
	84	struct blk_io_trace t;
	85	int fd, ret;
	86
	87	fd = open(filename, O_RDONLY);
	88	if (fd < 0) {
	89	perror("open blktrace");
	90	return 0;
	91	}
	92
	93	ret = read(fd, &t, sizeof(t));
	94	close(fd);
	95
	96	if (ret < 0) {
	97	perror("read blktrace");
	98	return 0;
	99	} else if (ret != sizeof(t)) {
	100	log_err("fio: short read on blktrace file\n");
	101	return 0;
	102	}
	103
	104	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
	105	return 1;
	106
	107	return 0;
	108	}
	109
8c1fdf04 JA	110	/*
	111	* Store blk_io_trace data in an ipo for later retrieval.
	112	*/
fdefd987	113	static void store_ipo(struct thread_data *td, unsigned long long offset,
8c1fdf04	114	unsigned int bytes, int rw, unsigned long long ttime)
fdefd987 JA	115	{
	116	struct io_piece ipo = malloc(sizeof(ipo));
	117
	118	memset(ipo, 0, sizeof(*ipo));
	119	INIT_LIST_HEAD(&ipo->list);
a2eea81b JA	120	/*
	121	* the 512 is wrong here, it should be the hardware sector size...
	122	*/
	123	ipo->offset = offset * 512;
fdefd987	124	ipo->len = bytes;
8c1fdf04	125	ipo->delay = ttime / 1000;
fdefd987 JA	126	if (rw)
	127	ipo->ddir = DDIR_WRITE;
	128	else
	129	ipo->ddir = DDIR_READ;
	130
	131	list_add_tail(&ipo->list, &td->io_log_list);
	132	}
	133
8c1fdf04 JA	134	/*
	135	* We only care for queue traces, most of the others are side effects
	136	* due to internal workings of the block layer.
	137	*/
	138	static void handle_trace(struct thread_data td, struct blk_io_trace t,
d84f8d49 JA	139	unsigned long long ttime, unsigned long *ios,
d84f8d49 JA	140	unsigned int *bs)
fb7b71a3	141	{
fdefd987 JA	142	int rw;
	143
	144	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
	145	return;
a2eea81b JA	146	if (t->action & BLK_TC_ACT(BLK_TC_PC))
	147	return;
	148
	149	/*
	150	* should not happen, need to look into that...
	151	*/
	152	if (!t->bytes)
	153	return;
fdefd987	154
e7a7d70b	155	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
d84f8d49 JA	156
	157	if (t->bytes > bs[rw])
	158	bs[rw] = t->bytes;
	159
8c1fdf04	160	ios[rw]++;
6df8adaa	161	td->o.size += t->bytes;
8c1fdf04	162	store_ipo(td, t->sector, t->bytes, rw, ttime);
fb7b71a3 JA	163	}
fb7b71a3 JA	164
8c1fdf04 JA	165	/*
	166	* Load a blktrace file by reading all the blk_io_trace entries, and storing
	167	* them as io_pieces like the fio text version would do.
	168	*/
fb7b71a3 JA	169	int load_blktrace(struct thread_data td, const char filename)
fb7b71a3 JA	170	{
a61eddec	171	unsigned long long ttime, delay;
fb7b71a3	172	struct blk_io_trace t;
8c1fdf04	173	unsigned long ios[2];
a61eddec	174	unsigned int cpu;
d84f8d49	175	unsigned int rw_bs[2];
e2887563	176	struct fifo *fifo;
fb7b71a3 JA	177	int fd;
	178
	179	fd = open(filename, O_RDONLY);
	180	if (fd < 0) {
	181	td_verror(td, errno, "open blktrace file");
	182	return 1;
	183	}
	184
e2887563 JA	185	fifo = fifo_alloc(TRACE_FIFO_SIZE);
e2887563 JA	186
6df8adaa JA	187	td->o.size = 0;
6df8adaa JA	188
a61eddec	189	cpu = 0;
d84f8d49 JA	190	ttime = 0;
	191	ios[0] = ios[1] = 0;
	192	rw_bs[0] = rw_bs[1] = 0;
fb7b71a3	193	do {
e2887563	194	int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
fb7b71a3	195
e2887563	196	if (ret < 0)
8c1fdf04	197	goto err;
e2887563 JA	198	else if (!ret)
	199	break;
	200	else if (ret < (int) sizeof(t)) {
	201	log_err("fio: short fifo get\n");
fb7b71a3	202	break;
fb7b71a3 JA	203	}
	204
	205	if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
	206	log_err("fio: bad magic in blktrace data\n");
8c1fdf04	207	goto err;
fb7b71a3 JA	208	}
	209	if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
	210	log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
8c1fdf04	211	goto err;
fb7b71a3 JA	212	}
	213	ret = discard_pdu(fd, &t);
	214	if (ret) {
	215	td_verror(td, ret, "blktrace lseek");
8c1fdf04	216	goto err;
fb7b71a3	217	}
a61eddec	218	if (!ttime) {
8c1fdf04	219	ttime = t.time;
a61eddec JA	220	cpu = t.cpu;
	221	}
	222	delay = 0;
	223	if (cpu == t.cpu)
	224	delay = t.time - ttime;
d84f8d49	225	handle_trace(td, &t, delay, ios, rw_bs);
8c1fdf04	226	ttime = t.time;
a61eddec	227	cpu = t.cpu;
fb7b71a3 JA	228	} while (1);
fb7b71a3 JA	229
38470f85	230	fifo_free(fifo);
fb7b71a3	231	close(fd);
8c1fdf04 JA	232
	233	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	234	log_err("fio: found no ios in blktrace data\n");
	235	return 1;
d84f8d49	236	} else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
8c1fdf04	237	td->o.td_ddir = TD_DDIR_READ;
d84f8d49 JA	238	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
d84f8d49 JA	239	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
8c1fdf04	240	td->o.td_ddir = TD_DDIR_WRITE;
d84f8d49 JA	241	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
d84f8d49 JA	242	} else {
8c1fdf04	243	td->o.td_ddir = TD_DDIR_RW;
d84f8d49 JA	244	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	245	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	246	}
8c1fdf04 JA	247
	248	/*
	249	* We need to do direct/raw ios to the device, to avoid getting
	250	* read-ahead in our way.
	251	*/
	252	td->o.odirect = 1;
	253
fb7b71a3	254	return 0;
8c1fdf04 JA	255	err:
8c1fdf04 JA	256	close(fd);
38470f85	257	fifo_free(fifo);
8c1fdf04	258	return 1;
fb7b71a3	259	}