[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)
{
	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(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", 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);
		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:
		printf("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);
		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);
}

/*
 * 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, i;
	struct fio_file *f;

	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)) == 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);

	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
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	*/
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
d1c46c04 DN	100	static int lookup_device(struct thread_data td, char path, unsigned int maj,
d1c46c04 DN	101	unsigned int min)
5e6c2067 JA	102	{
	103	struct dirent *dir;
	104	struct stat st;
	105	int found = 0;
	106	DIR *D;
	107
	108	D = opendir(path);
	109	if (!D)
	110	return 0;
	111
	112	while ((dir = readdir(D)) != NULL) {
	113	char full_path[256];
	114
	115	if (!strcmp(dir->d_name, ".") \|\| !strcmp(dir->d_name, ".."))
	116	continue;
	117
b9fd788f	118	sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name);
5e6c2067 JA	119	if (lstat(full_path, &st) == -1) {
	120	perror("lstat");
	121	break;
	122	}
	123
	124	if (S_ISDIR(st.st_mode)) {
d1c46c04	125	found = lookup_device(td, full_path, maj, min);
5e6c2067 JA	126	if (found) {
	127	strcpy(path, full_path);
	128	break;
	129	}
	130	}
	131
	132	if (!S_ISBLK(st.st_mode))
	133	continue;
	134
d1c46c04 DN	135	/*
	136	* If replay_redirect is set then always return this device
	137	* upon lookup which overrides the device lookup based on
	138	* major minor in the actual blktrace
	139	*/
	140	if (td->o.replay_redirect) {
	141	dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
	142	" with: %s", maj, min,
	143	td->o.replay_redirect);
	144	strcpy(path, td->o.replay_redirect);
	145	found = 1;
	146	break;
	147	}
	148
5e6c2067	149	if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
5ec10eaa	150	dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
5e6c2067 JA	151	strcpy(path, full_path);
	152	found = 1;
	153	break;
	154	}
	155	}
	156
	157	closedir(D);
	158	return found;
	159	}
	160
c69aa91f JA	161	#define FMINORBITS 20
	162	#define FMINORMASK ((1U << FMINORBITS) - 1)
	163	#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
	164	#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
eeb9c2aa	165
89ac1d48	166	static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
691c8fb0 JA	167	{
	168	struct io_piece *ipo;
	169
	170	ipo = calloc(1, sizeof(*ipo));
0d29de83	171	init_ipo(ipo);
691c8fb0 JA	172
	173	ipo->ddir = DDIR_INVAL;
	174	ipo->fileno = fileno;
89ac1d48	175	ipo->file_action = action;
01743ee1	176	flist_add_tail(&ipo->list, &td->io_log_list);
691c8fb0 JA	177	}
691c8fb0 JA	178
89ac1d48	179	static int trace_add_file(struct thread_data *td, __u32 device)
5e6c2067	180	{
89ac1d48	181	static unsigned int last_maj, last_min, last_fileno;
c69aa91f JA	182	unsigned int maj = FMAJOR(device);
c69aa91f JA	183	unsigned int min = FMINOR(device);
5e6c2067 JA	184	struct fio_file *f;
	185	char dev[256];
	186	unsigned int i;
	187
	188	if (last_maj == maj && last_min == min)
89ac1d48	189	return last_fileno;
5e6c2067 JA	190
	191	last_maj = maj;
	192	last_min = min;
	193
	194	/*
	195	* check for this file in our list
	196	*/
	197	for_each_file(td, f, i)
89ac1d48 SL	198	if (f->major == maj && f->minor == min) {
	199	last_fileno = f->fileno;
	200	return last_fileno;
	201	}
5e6c2067 JA	202
5e6c2067 JA	203	strcpy(dev, "/dev");
d1c46c04	204	if (lookup_device(td, dev, maj, min)) {
691c8fb0 JA	205	int fileno;
691c8fb0 JA	206
bd6f78b2	207	dprint(FD_BLKTRACE, "add devices %s\n", dev);
49ffb4a2	208	fileno = add_file_exclusive(td, dev);
89ac1d48 SL	209	trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
89ac1d48 SL	210	last_fileno = fileno;
bd6f78b2	211	}
89ac1d48	212	return last_fileno;
5e6c2067 JA	213	}
5e6c2067 JA	214
8c1fdf04 JA	215	/*
	216	* Store blk_io_trace data in an ipo for later retrieval.
	217	*/
fdefd987	218	static void store_ipo(struct thread_data *td, unsigned long long offset,
89ac1d48 SL	219	unsigned int bytes, int rw, unsigned long long ttime,
89ac1d48 SL	220	int fileno)
fdefd987 JA	221	{
	222	struct io_piece ipo = malloc(sizeof(ipo));
	223
0d29de83 JA	224	init_ipo(ipo);
0d29de83 JA	225
a2eea81b JA	226	/*
	227	* the 512 is wrong here, it should be the hardware sector size...
	228	*/
	229	ipo->offset = offset * 512;
fdefd987	230	ipo->len = bytes;
8c1fdf04	231	ipo->delay = ttime / 1000;
fdefd987 JA	232	if (rw)
	233	ipo->ddir = DDIR_WRITE;
	234	else
	235	ipo->ddir = DDIR_READ;
89ac1d48	236	ipo->fileno = fileno;
fdefd987	237
bd6f78b2 JA	238	dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
	239	ipo->ddir, ipo->offset,
	240	ipo->len, ipo->delay);
691c8fb0	241	queue_io_piece(td, ipo);
fdefd987 JA	242	}
fdefd987 JA	243
0b9d69ec	244	static void handle_trace_notify(struct blk_io_trace *t)
cd991b9e	245	{
691c8fb0 JA	246	switch (t->action) {
	247	case BLK_TN_PROCESS:
	248	printf("got process notify: %x, %d\n", t->action, t->pid);
	249	break;
	250	case BLK_TN_TIMESTAMP:
	251	printf("got timestamp notify: %x, %d\n", t->action, t->pid);
	252	break;
ff58fced JA	253	case BLK_TN_MESSAGE:
ff58fced JA	254	break;
691c8fb0 JA	255	default:
	256	dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
	257	break;
	258	}
	259	}
5b3023b8	260
ff58fced JA	261	static void handle_trace_discard(struct thread_data td, struct blk_io_trace t,
	262	unsigned long long ttime, unsigned long *ios)
	263	{
	264	struct io_piece ipo = malloc(sizeof(ipo));
89ac1d48	265	int fileno;
ff58fced	266
0d29de83	267	init_ipo(ipo);
89ac1d48	268	fileno = trace_add_file(td, t->device);
ff58fced JA	269
	270	ios[DDIR_WRITE]++;
	271	td->o.size += t->bytes;
	272
	273	memset(ipo, 0, sizeof(*ipo));
	274	INIT_FLIST_HEAD(&ipo->list);
	275
	276	/*
	277	* the 512 is wrong here, it should be the hardware sector size...
	278	*/
	279	ipo->offset = t->sector * 512;
	280	ipo->len = t->bytes;
	281	ipo->delay = ttime / 1000;
	282	ipo->ddir = DDIR_TRIM;
89ac1d48	283	ipo->fileno = fileno;
ff58fced JA	284
	285	dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
	286	ipo->offset, ipo->len,
	287	ipo->delay);
	288	queue_io_piece(td, ipo);
	289	}
	290
691c8fb0 JA	291	static void handle_trace_fs(struct thread_data td, struct blk_io_trace t,
	292	unsigned long long ttime, unsigned long *ios,
	293	unsigned int *bs)
	294	{
	295	int rw;
89ac1d48	296	int fileno;
5b3023b8	297
89ac1d48	298	fileno = trace_add_file(td, t->device);
5b3023b8 JA	299
	300	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
	301
	302	if (t->bytes > bs[rw])
	303	bs[rw] = t->bytes;
	304
	305	ios[rw]++;
	306	td->o.size += t->bytes;
89ac1d48	307	store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
cd991b9e JA	308	}
cd991b9e JA	309
691c8fb0 JA	310	/*
	311	* We only care for queue traces, most of the others are side effects
	312	* due to internal workings of the block layer.
	313	*/
	314	static void handle_trace(struct thread_data td, struct blk_io_trace t,
	315	unsigned long long ttime, unsigned long *ios,
	316	unsigned int *bs)
	317	{
	318	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
	319	return;
	320	if (t->action & BLK_TC_ACT(BLK_TC_PC))
	321	return;
	322
	323	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
aec2de20	324	handle_trace_notify(t);
ff58fced JA	325	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
ff58fced JA	326	handle_trace_discard(td, t, ttime, ios);
691c8fb0 JA	327	else
	328	handle_trace_fs(td, t, ttime, ios, bs);
	329	}
	330
8c1fdf04 JA	331	/*
	332	* Load a blktrace file by reading all the blk_io_trace entries, and storing
	333	* them as io_pieces like the fio text version would do.
	334	*/
fb7b71a3 JA	335	int load_blktrace(struct thread_data td, const char filename)
fb7b71a3 JA	336	{
a61eddec	337	unsigned long long ttime, delay;
fb7b71a3	338	struct blk_io_trace t;
4241ea8f	339	unsigned long ios[2], skipped_writes;
a61eddec	340	unsigned int cpu;
d84f8d49	341	unsigned int rw_bs[2];
e2887563	342	struct fifo *fifo;
89ac1d48 SL	343	int fd, i;
89ac1d48 SL	344	struct fio_file *f;
fb7b71a3 JA	345
	346	fd = open(filename, O_RDONLY);
	347	if (fd < 0) {
	348	td_verror(td, errno, "open blktrace file");
	349	return 1;
	350	}
	351
e2887563 JA	352	fifo = fifo_alloc(TRACE_FIFO_SIZE);
e2887563 JA	353
6df8adaa JA	354	td->o.size = 0;
6df8adaa JA	355
a61eddec	356	cpu = 0;
d84f8d49 JA	357	ttime = 0;
	358	ios[0] = ios[1] = 0;
	359	rw_bs[0] = rw_bs[1] = 0;
4241ea8f	360	skipped_writes = 0;
fb7b71a3	361	do {
e2887563	362	int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
fb7b71a3	363
e2887563	364	if (ret < 0)
8c1fdf04	365	goto err;
e2887563 JA	366	else if (!ret)
	367	break;
	368	else if (ret < (int) sizeof(t)) {
	369	log_err("fio: short fifo get\n");
fb7b71a3	370	break;
fb7b71a3 JA	371	}
	372
	373	if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
5ec10eaa JA	374	log_err("fio: bad magic in blktrace data: %x\n",
5ec10eaa JA	375	t.magic);
8c1fdf04	376	goto err;
fb7b71a3 JA	377	}
fb7b71a3 JA	378	if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
5ec10eaa JA	379	log_err("fio: bad blktrace version %d\n",
5ec10eaa JA	380	t.magic & 0xff);
8c1fdf04	381	goto err;
fb7b71a3	382	}
f12b323f JA	383	ret = discard_pdu(td, fifo, fd, &t);
f12b323f JA	384	if (ret < 0) {
fb7b71a3	385	td_verror(td, ret, "blktrace lseek");
8c1fdf04	386	goto err;
f12b323f JA	387	} else if (t.pdu_len != ret) {
	388	log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
	389	goto err;
fb7b71a3	390	}
691c8fb0 JA	391	if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
	392	if (!ttime) {
	393	ttime = t.time;
	394	cpu = t.cpu;
	395	}
	396
	397	delay = 0;
	398	if (cpu == t.cpu)
	399	delay = t.time - ttime;
	400	if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
	401	skipped_writes++;
64bbb865 DN	402	else {
	403	/*
	404	* set delay to zero if no_stall enabled for
	405	* fast replay
	406	*/
	407	if (td->o.no_stall)
	408	delay = 0;
	409
691c8fb0	410	handle_trace(td, &t, delay, ios, rw_bs);
64bbb865	411	}
691c8fb0	412
8c1fdf04	413	ttime = t.time;
a61eddec	414	cpu = t.cpu;
a6edd638 JA	415	} else {
a6edd638 JA	416	delay = 0;
4241ea8f	417	handle_trace(td, &t, delay, ios, rw_bs);
a6edd638	418	}
fb7b71a3 JA	419	} while (1);
fb7b71a3 JA	420
85a47ca2 JA	421	for (i = 0; i < td->files_index; i++) {
85a47ca2 JA	422	f= td->files[i];
89ac1d48	423	trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
85a47ca2	424	}
89ac1d48	425
38470f85	426	fifo_free(fifo);
fb7b71a3	427	close(fd);
8c1fdf04	428
4241ea8f	429	if (skipped_writes)
5ec10eaa JA	430	log_err("fio: %s skips replay of %lu writes due to read-only\n",
5ec10eaa JA	431	td->o.name, skipped_writes);
4241ea8f	432
8c1fdf04 JA	433	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	434	log_err("fio: found no ios in blktrace data\n");
	435	return 1;
d84f8d49	436	} else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
8c1fdf04	437	td->o.td_ddir = TD_DDIR_READ;
d84f8d49 JA	438	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
d84f8d49 JA	439	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
8c1fdf04	440	td->o.td_ddir = TD_DDIR_WRITE;
d84f8d49 JA	441	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
d84f8d49 JA	442	} else {
8c1fdf04	443	td->o.td_ddir = TD_DDIR_RW;
d84f8d49 JA	444	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	445	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	446	}
8c1fdf04 JA	447
	448	/*
	449	* We need to do direct/raw ios to the device, to avoid getting
	450	* read-ahead in our way.
	451	*/
	452	td->o.odirect = 1;
	453
fb7b71a3	454	return 0;
8c1fdf04 JA	455	err:
8c1fdf04 JA	456	close(fd);
38470f85	457	fifo_free(fifo);
8c1fdf04	458	return 1;
fb7b71a3	459	}