[fio.git] / blktrace.c

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

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

#define TRACE_FIFO_SIZE	65536

/*
 * 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);

	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;

	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)
{
	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;
}

static int lookup_device(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", path, dir->d_name);
		if (lstat(full_path, &st) == -1) {
			perror("lstat");
			break;
		}

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

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

		if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
			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_file(struct thread_data *td, __u32 device)
{
	static unsigned int last_maj, last_min;
	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_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;

	strcpy(dev, "/dev");
	if (lookup_device(dev, maj, min))
		add_file(td, dev);
}

/*
 * 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;
	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
		return;

	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);
}

/*
 * 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: %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 (!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>
5e6c2067 JA	6	#include <sys/stat.h>
5e6c2067 JA	7	#include <dirent.h>
8c1fdf04	8
fb7b71a3 JA	9	#include "list.h"
	10	#include "fio.h"
	11	#include "blktrace_api.h"
	12
f12b323f	13	#define TRACE_FIFO_SIZE 65536
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
	37	return ret;
e2887563 JA	38	}
	39
	40	/*
	41	* Retrieve 'len' bytes from the fifo, refilling if necessary.
	42	*/
	43	static int trace_fifo_get(struct thread_data td, struct fifo fifo, int fd,
	44	void *buf, unsigned int len)
	45	{
f12b323f JA	46	if (fifo_len(fifo) < len) {
f12b323f JA	47	int ret = refill_fifo(td, fifo, fd);
e2887563	48
f12b323f JA	49	if (ret < 0)
	50	return ret;
	51	}
e2887563 JA	52
	53	return fifo_get(fifo, buf, len);
	54	}
	55
8c1fdf04 JA	56	/*
	57	* Just discard the pdu by seeking past it.
	58	*/
f12b323f JA	59	static int discard_pdu(struct thread_data td, struct fifo fifo, int fd,
f12b323f JA	60	struct blk_io_trace *t)
fb7b71a3 JA	61	{
	62	if (t->pdu_len == 0)
	63	return 0;
	64
f12b323f	65	return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
fb7b71a3 JA	66	}
fb7b71a3 JA	67
8c1fdf04 JA	68	/*
	69	* Check if this is a blktrace binary data file. We read a single trace
	70	* into memory and check for the magic signature.
	71	*/
fb7b71a3 JA	72	int is_blktrace(const char *filename)
	73	{
	74	struct blk_io_trace t;
	75	int fd, ret;
	76
	77	fd = open(filename, O_RDONLY);
	78	if (fd < 0) {
	79	perror("open blktrace");
	80	return 0;
	81	}
	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
	94	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
	95	return 1;
	96
	97	return 0;
	98	}
	99
5e6c2067 JA	100	static int lookup_device(char *path, unsigned int maj, unsigned int min)
	101	{
	102	struct dirent *dir;
	103	struct stat st;
	104	int found = 0;
	105	DIR *D;
	106
	107	D = opendir(path);
	108	if (!D)
	109	return 0;
	110
	111	while ((dir = readdir(D)) != NULL) {
	112	char full_path[256];
	113
	114	if (!strcmp(dir->d_name, ".") \|\| !strcmp(dir->d_name, ".."))
	115	continue;
	116
	117	sprintf(full_path, "%s/%s", path, dir->d_name);
	118	if (lstat(full_path, &st) == -1) {
	119	perror("lstat");
	120	break;
	121	}
	122
	123	if (S_ISDIR(st.st_mode)) {
	124	found = lookup_device(full_path, maj, min);
	125	if (found) {
	126	strcpy(path, full_path);
	127	break;
	128	}
	129	}
	130
	131	if (!S_ISBLK(st.st_mode))
	132	continue;
	133
	134	if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
	135	strcpy(path, full_path);
	136	found = 1;
	137	break;
	138	}
	139	}
	140
	141	closedir(D);
	142	return found;
	143	}
	144
c69aa91f JA	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))
eeb9c2aa	149
5e6c2067 JA	150	static void trace_add_file(struct thread_data *td, __u32 device)
	151	{
	152	static unsigned int last_maj, last_min;
c69aa91f JA	153	unsigned int maj = FMAJOR(device);
c69aa91f JA	154	unsigned int min = FMINOR(device);
5e6c2067 JA	155	struct fio_file *f;
	156	char dev[256];
	157	unsigned int i;
	158
	159	if (last_maj == maj && last_min == min)
	160	return;
	161
	162	last_maj = maj;
	163	last_min = min;
	164
	165	/*
	166	* check for this file in our list
	167	*/
	168	for_each_file(td, f, i)
	169	if (f->major == maj && f->minor == min)
	170	return;
	171
	172	strcpy(dev, "/dev");
	173	if (lookup_device(dev, maj, min))
	174	add_file(td, dev);
	175	}
	176
8c1fdf04 JA	177	/*
	178	* Store blk_io_trace data in an ipo for later retrieval.
	179	*/
fdefd987	180	static void store_ipo(struct thread_data *td, unsigned long long offset,
8c1fdf04	181	unsigned int bytes, int rw, unsigned long long ttime)
fdefd987 JA	182	{
	183	struct io_piece ipo = malloc(sizeof(ipo));
	184
	185	memset(ipo, 0, sizeof(*ipo));
	186	INIT_LIST_HEAD(&ipo->list);
a2eea81b JA	187	/*
	188	* the 512 is wrong here, it should be the hardware sector size...
	189	*/
	190	ipo->offset = offset * 512;
fdefd987	191	ipo->len = bytes;
8c1fdf04	192	ipo->delay = ttime / 1000;
fdefd987 JA	193	if (rw)
	194	ipo->ddir = DDIR_WRITE;
	195	else
	196	ipo->ddir = DDIR_READ;
	197
	198	list_add_tail(&ipo->list, &td->io_log_list);
	199	}
	200
8c1fdf04 JA	201	/*
	202	* We only care for queue traces, most of the others are side effects
	203	* due to internal workings of the block layer.
	204	*/
	205	static void handle_trace(struct thread_data td, struct blk_io_trace t,
d84f8d49 JA	206	unsigned long long ttime, unsigned long *ios,
d84f8d49 JA	207	unsigned int *bs)
fb7b71a3	208	{
fdefd987 JA	209	int rw;
	210
	211	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
	212	return;
a2eea81b JA	213	if (t->action & BLK_TC_ACT(BLK_TC_PC))
a2eea81b JA	214	return;
3f3d3616	215	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
a2eea81b	216	return;
fdefd987	217
5e6c2067 JA	218	trace_add_file(td, t->device);
5e6c2067 JA	219
e7a7d70b	220	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
d84f8d49 JA	221
	222	if (t->bytes > bs[rw])
	223	bs[rw] = t->bytes;
	224
8c1fdf04	225	ios[rw]++;
6df8adaa	226	td->o.size += t->bytes;
8c1fdf04	227	store_ipo(td, t->sector, t->bytes, rw, ttime);
fb7b71a3 JA	228	}
fb7b71a3 JA	229
8c1fdf04 JA	230	/*
	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.
	233	*/
fb7b71a3 JA	234	int load_blktrace(struct thread_data td, const char filename)
fb7b71a3 JA	235	{
a61eddec	236	unsigned long long ttime, delay;
fb7b71a3	237	struct blk_io_trace t;
8c1fdf04	238	unsigned long ios[2];
a61eddec	239	unsigned int cpu;
d84f8d49	240	unsigned int rw_bs[2];
e2887563	241	struct fifo *fifo;
fb7b71a3 JA	242	int fd;
	243
	244	fd = open(filename, O_RDONLY);
	245	if (fd < 0) {
	246	td_verror(td, errno, "open blktrace file");
	247	return 1;
	248	}
	249
e2887563 JA	250	fifo = fifo_alloc(TRACE_FIFO_SIZE);
e2887563 JA	251
6df8adaa JA	252	td->o.size = 0;
6df8adaa JA	253
a61eddec	254	cpu = 0;
d84f8d49 JA	255	ttime = 0;
	256	ios[0] = ios[1] = 0;
	257	rw_bs[0] = rw_bs[1] = 0;
fb7b71a3	258	do {
e2887563	259	int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
fb7b71a3	260
e2887563	261	if (ret < 0)
8c1fdf04	262	goto err;
e2887563 JA	263	else if (!ret)
	264	break;
	265	else if (ret < (int) sizeof(t)) {
	266	log_err("fio: short fifo get\n");
fb7b71a3	267	break;
fb7b71a3 JA	268	}
	269
	270	if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
f12b323f	271	log_err("fio: bad magic in blktrace data: %x\n", t.magic);
8c1fdf04	272	goto err;
fb7b71a3 JA	273	}
	274	if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
	275	log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
8c1fdf04	276	goto err;
fb7b71a3	277	}
f12b323f JA	278	ret = discard_pdu(td, fifo, fd, &t);
f12b323f JA	279	if (ret < 0) {
fb7b71a3	280	td_verror(td, ret, "blktrace lseek");
8c1fdf04	281	goto err;
f12b323f JA	282	} else if (t.pdu_len != ret) {
	283	log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
	284	goto err;
fb7b71a3	285	}
a61eddec	286	if (!ttime) {
8c1fdf04	287	ttime = t.time;
a61eddec JA	288	cpu = t.cpu;
	289	}
	290	delay = 0;
	291	if (cpu == t.cpu)
	292	delay = t.time - ttime;
d84f8d49	293	handle_trace(td, &t, delay, ios, rw_bs);
8c1fdf04	294	ttime = t.time;
a61eddec	295	cpu = t.cpu;
fb7b71a3 JA	296	} while (1);
fb7b71a3 JA	297
38470f85	298	fifo_free(fifo);
fb7b71a3	299	close(fd);
8c1fdf04 JA	300
	301	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	302	log_err("fio: found no ios in blktrace data\n");
	303	return 1;
d84f8d49	304	} else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
8c1fdf04	305	td->o.td_ddir = TD_DDIR_READ;
d84f8d49 JA	306	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
d84f8d49 JA	307	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
8c1fdf04	308	td->o.td_ddir = TD_DDIR_WRITE;
d84f8d49 JA	309	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
d84f8d49 JA	310	} else {
8c1fdf04	311	td->o.td_ddir = TD_DDIR_RW;
d84f8d49 JA	312	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	313	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	314	}
8c1fdf04 JA	315
	316	/*
	317	* We need to do direct/raw ios to the device, to avoid getting
	318	* read-ahead in our way.
	319	*/
	320	td->o.odirect = 1;
	321
fb7b71a3	322	return 0;
8c1fdf04 JA	323	err:
8c1fdf04 JA	324	close(fd);
38470f85	325	fifo_free(fifo);
8c1fdf04	326	return 1;
fb7b71a3	327	}