[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);

	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)
{
	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)
		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)) {
			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_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)) {
		dprint(FD_BLKTRACE, "add devices %s\n", dev);
		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;

	dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
							ipo->ddir, ipo->offset,
							ipo->len, ipo->delay);
	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], skipped_writes;
	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;
	skipped_writes = 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 (t.action & BLK_TC_ACT(BLK_TC_NOTIFY))
			continue;
		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
			handle_trace(td, &t, delay, ios, rw_bs);
		ttime = t.time;
		cpu = t.cpu;
	} while (1);

	fifo_free(fifo);
	close(fd);

	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;

	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
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	*/
fb7b71a3 JA	74	int is_blktrace(const char *filename)
	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
	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)) {
bd6f78b2	135	dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj,min);
5e6c2067 JA	136	strcpy(path, full_path);
	137	found = 1;
	138	break;
	139	}
	140	}
	141
	142	closedir(D);
	143	return found;
	144	}
	145
c69aa91f JA	146	#define FMINORBITS 20
	147	#define FMINORMASK ((1U << FMINORBITS) - 1)
	148	#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
	149	#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
eeb9c2aa	150
5e6c2067 JA	151	static void trace_add_file(struct thread_data *td, __u32 device)
	152	{
	153	static unsigned int last_maj, last_min;
c69aa91f JA	154	unsigned int maj = FMAJOR(device);
c69aa91f JA	155	unsigned int min = FMINOR(device);
5e6c2067 JA	156	struct fio_file *f;
	157	char dev[256];
	158	unsigned int i;
	159
	160	if (last_maj == maj && last_min == min)
	161	return;
	162
	163	last_maj = maj;
	164	last_min = min;
	165
	166	/*
	167	* check for this file in our list
	168	*/
	169	for_each_file(td, f, i)
	170	if (f->major == maj && f->minor == min)
	171	return;
	172
	173	strcpy(dev, "/dev");
bd6f78b2 JA	174	if (lookup_device(dev, maj, min)) {
bd6f78b2 JA	175	dprint(FD_BLKTRACE, "add devices %s\n", dev);
5e6c2067	176	add_file(td, dev);
bd6f78b2	177	}
5e6c2067 JA	178	}
5e6c2067 JA	179
8c1fdf04 JA	180	/*
	181	* Store blk_io_trace data in an ipo for later retrieval.
	182	*/
fdefd987	183	static void store_ipo(struct thread_data *td, unsigned long long offset,
8c1fdf04	184	unsigned int bytes, int rw, unsigned long long ttime)
fdefd987 JA	185	{
	186	struct io_piece ipo = malloc(sizeof(ipo));
	187
	188	memset(ipo, 0, sizeof(*ipo));
	189	INIT_LIST_HEAD(&ipo->list);
a2eea81b JA	190	/*
	191	* the 512 is wrong here, it should be the hardware sector size...
	192	*/
	193	ipo->offset = offset * 512;
fdefd987	194	ipo->len = bytes;
8c1fdf04	195	ipo->delay = ttime / 1000;
fdefd987 JA	196	if (rw)
	197	ipo->ddir = DDIR_WRITE;
	198	else
	199	ipo->ddir = DDIR_READ;
	200
bd6f78b2 JA	201	dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
	202	ipo->ddir, ipo->offset,
	203	ipo->len, ipo->delay);
fdefd987 JA	204	list_add_tail(&ipo->list, &td->io_log_list);
	205	}
	206
8c1fdf04 JA	207	/*
	208	* We only care for queue traces, most of the others are side effects
	209	* due to internal workings of the block layer.
	210	*/
	211	static void handle_trace(struct thread_data td, struct blk_io_trace t,
d84f8d49 JA	212	unsigned long long ttime, unsigned long *ios,
d84f8d49 JA	213	unsigned int *bs)
fb7b71a3	214	{
fdefd987 JA	215	int rw;
	216
	217	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
	218	return;
a2eea81b JA	219	if (t->action & BLK_TC_ACT(BLK_TC_PC))
a2eea81b JA	220	return;
3f3d3616	221	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
a2eea81b	222	return;
fdefd987	223
5e6c2067 JA	224	trace_add_file(td, t->device);
5e6c2067 JA	225
e7a7d70b	226	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
d84f8d49 JA	227
	228	if (t->bytes > bs[rw])
	229	bs[rw] = t->bytes;
	230
8c1fdf04	231	ios[rw]++;
6df8adaa	232	td->o.size += t->bytes;
8c1fdf04	233	store_ipo(td, t->sector, t->bytes, rw, ttime);
fb7b71a3 JA	234	}
fb7b71a3 JA	235
8c1fdf04 JA	236	/*
	237	* Load a blktrace file by reading all the blk_io_trace entries, and storing
	238	* them as io_pieces like the fio text version would do.
	239	*/
fb7b71a3 JA	240	int load_blktrace(struct thread_data td, const char filename)
fb7b71a3 JA	241	{
a61eddec	242	unsigned long long ttime, delay;
fb7b71a3	243	struct blk_io_trace t;
4241ea8f	244	unsigned long ios[2], skipped_writes;
a61eddec	245	unsigned int cpu;
d84f8d49	246	unsigned int rw_bs[2];
e2887563	247	struct fifo *fifo;
fb7b71a3 JA	248	int fd;
	249
	250	fd = open(filename, O_RDONLY);
	251	if (fd < 0) {
	252	td_verror(td, errno, "open blktrace file");
	253	return 1;
	254	}
	255
e2887563 JA	256	fifo = fifo_alloc(TRACE_FIFO_SIZE);
e2887563 JA	257
6df8adaa JA	258	td->o.size = 0;
6df8adaa JA	259
a61eddec	260	cpu = 0;
d84f8d49 JA	261	ttime = 0;
	262	ios[0] = ios[1] = 0;
	263	rw_bs[0] = rw_bs[1] = 0;
4241ea8f	264	skipped_writes = 0;
fb7b71a3	265	do {
e2887563	266	int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
fb7b71a3	267
e2887563	268	if (ret < 0)
8c1fdf04	269	goto err;
e2887563 JA	270	else if (!ret)
	271	break;
	272	else if (ret < (int) sizeof(t)) {
	273	log_err("fio: short fifo get\n");
fb7b71a3	274	break;
fb7b71a3 JA	275	}
	276
	277	if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
f12b323f	278	log_err("fio: bad magic in blktrace data: %x\n", t.magic);
8c1fdf04	279	goto err;
fb7b71a3 JA	280	}
	281	if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
	282	log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
8c1fdf04	283	goto err;
fb7b71a3	284	}
f12b323f JA	285	ret = discard_pdu(td, fifo, fd, &t);
f12b323f JA	286	if (ret < 0) {
fb7b71a3	287	td_verror(td, ret, "blktrace lseek");
8c1fdf04	288	goto err;
f12b323f JA	289	} else if (t.pdu_len != ret) {
	290	log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
	291	goto err;
fb7b71a3	292	}
f3f72c5e JA	293	if (t.action & BLK_TC_ACT(BLK_TC_NOTIFY))
f3f72c5e JA	294	continue;
a61eddec	295	if (!ttime) {
8c1fdf04	296	ttime = t.time;
a61eddec JA	297	cpu = t.cpu;
	298	}
	299	delay = 0;
	300	if (cpu == t.cpu)
	301	delay = t.time - ttime;
4241ea8f JA	302	if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
	303	skipped_writes++;
	304	else
	305	handle_trace(td, &t, delay, ios, rw_bs);
8c1fdf04	306	ttime = t.time;
a61eddec	307	cpu = t.cpu;
fb7b71a3 JA	308	} while (1);
fb7b71a3 JA	309
38470f85	310	fifo_free(fifo);
fb7b71a3	311	close(fd);
8c1fdf04	312
4241ea8f JA	313	if (skipped_writes)
	314	log_err("fio: <%s> skips replay of %lu writes due to read-only\n", td->o.name, skipped_writes);
	315
8c1fdf04 JA	316	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	317	log_err("fio: found no ios in blktrace data\n");
	318	return 1;
d84f8d49	319	} else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
8c1fdf04	320	td->o.td_ddir = TD_DDIR_READ;
d84f8d49 JA	321	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
d84f8d49 JA	322	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
8c1fdf04	323	td->o.td_ddir = TD_DDIR_WRITE;
d84f8d49 JA	324	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
d84f8d49 JA	325	} else {
8c1fdf04	326	td->o.td_ddir = TD_DDIR_RW;
d84f8d49 JA	327	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	328	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	329	}
8c1fdf04 JA	330
	331	/*
	332	* We need to do direct/raw ios to the device, to avoid getting
	333	* read-ahead in our way.
	334	*/
	335	td->o.odirect = 1;
	336
fb7b71a3	337	return 0;
8c1fdf04 JA	338	err:
8c1fdf04 JA	339	close(fd);
38470f85	340	fifo_free(fifo);
8c1fdf04	341	return 1;
fb7b71a3	342	}