[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", 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:
		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;
	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 (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 (!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	*/
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"
	153	" with: %s", maj, min,
	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);
89ac1d48 SL	220	trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
89ac1d48 SL	221	last_fileno = fileno;
bd6f78b2	222	}
89ac1d48	223	return last_fileno;
5e6c2067 JA	224	}
5e6c2067 JA	225
8c1fdf04 JA	226	/*
	227	* Store blk_io_trace data in an ipo for later retrieval.
	228	*/
fdefd987	229	static void store_ipo(struct thread_data *td, unsigned long long offset,
89ac1d48 SL	230	unsigned int bytes, int rw, unsigned long long ttime,
89ac1d48 SL	231	int fileno)
fdefd987 JA	232	{
	233	struct io_piece ipo = malloc(sizeof(ipo));
	234
0d29de83 JA	235	init_ipo(ipo);
0d29de83 JA	236
a2eea81b JA	237	/*
	238	* the 512 is wrong here, it should be the hardware sector size...
	239	*/
	240	ipo->offset = offset * 512;
fdefd987	241	ipo->len = bytes;
8c1fdf04	242	ipo->delay = ttime / 1000;
fdefd987 JA	243	if (rw)
	244	ipo->ddir = DDIR_WRITE;
	245	else
	246	ipo->ddir = DDIR_READ;
89ac1d48	247	ipo->fileno = fileno;
fdefd987	248
bd6f78b2 JA	249	dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
	250	ipo->ddir, ipo->offset,
	251	ipo->len, ipo->delay);
691c8fb0	252	queue_io_piece(td, ipo);
fdefd987 JA	253	}
fdefd987 JA	254
0b9d69ec	255	static void handle_trace_notify(struct blk_io_trace *t)
cd991b9e	256	{
691c8fb0 JA	257	switch (t->action) {
691c8fb0 JA	258	case BLK_TN_PROCESS:
d95b34a6 JA	259	log_info("blktrace: got process notify: %x, %d\n",
d95b34a6 JA	260	t->action, t->pid);
691c8fb0 JA	261	break;
691c8fb0 JA	262	case BLK_TN_TIMESTAMP:
d95b34a6 JA	263	log_info("blktrace: got timestamp notify: %x, %d\n",
d95b34a6 JA	264	t->action, t->pid);
691c8fb0	265	break;
ff58fced JA	266	case BLK_TN_MESSAGE:
ff58fced JA	267	break;
691c8fb0 JA	268	default:
	269	dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
	270	break;
	271	}
	272	}
5b3023b8	273
ff58fced JA	274	static void handle_trace_discard(struct thread_data td, struct blk_io_trace t,
	275	unsigned long long ttime, unsigned long *ios)
	276	{
	277	struct io_piece ipo = malloc(sizeof(ipo));
89ac1d48	278	int fileno;
ff58fced	279
0d29de83	280	init_ipo(ipo);
89ac1d48	281	fileno = trace_add_file(td, t->device);
ff58fced JA	282
	283	ios[DDIR_WRITE]++;
	284	td->o.size += t->bytes;
	285
	286	memset(ipo, 0, sizeof(*ipo));
	287	INIT_FLIST_HEAD(&ipo->list);
	288
	289	/*
	290	* the 512 is wrong here, it should be the hardware sector size...
	291	*/
	292	ipo->offset = t->sector * 512;
	293	ipo->len = t->bytes;
	294	ipo->delay = ttime / 1000;
	295	ipo->ddir = DDIR_TRIM;
89ac1d48	296	ipo->fileno = fileno;
ff58fced JA	297
	298	dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
	299	ipo->offset, ipo->len,
	300	ipo->delay);
	301	queue_io_piece(td, ipo);
	302	}
	303
691c8fb0 JA	304	static void handle_trace_fs(struct thread_data td, struct blk_io_trace t,
	305	unsigned long long ttime, unsigned long *ios,
	306	unsigned int *bs)
	307	{
	308	int rw;
89ac1d48	309	int fileno;
5b3023b8	310
89ac1d48	311	fileno = trace_add_file(td, t->device);
5b3023b8 JA	312
	313	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
	314
	315	if (t->bytes > bs[rw])
	316	bs[rw] = t->bytes;
	317
	318	ios[rw]++;
	319	td->o.size += t->bytes;
89ac1d48	320	store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
cd991b9e JA	321	}
cd991b9e JA	322
691c8fb0 JA	323	/*
	324	* We only care for queue traces, most of the others are side effects
	325	* due to internal workings of the block layer.
	326	*/
	327	static void handle_trace(struct thread_data td, struct blk_io_trace t,
	328	unsigned long long ttime, unsigned long *ios,
	329	unsigned int *bs)
	330	{
	331	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
	332	return;
	333	if (t->action & BLK_TC_ACT(BLK_TC_PC))
	334	return;
	335
	336	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
aec2de20	337	handle_trace_notify(t);
ff58fced JA	338	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
ff58fced JA	339	handle_trace_discard(td, t, ttime, ios);
691c8fb0 JA	340	else
	341	handle_trace_fs(td, t, ttime, ios, bs);
	342	}
	343
d95b34a6 JA	344	static void byteswap_trace(struct blk_io_trace *t)
	345	{
	346	t->magic = fio_swap32(t->magic);
	347	t->sequence = fio_swap32(t->sequence);
	348	t->time = fio_swap64(t->time);
	349	t->sector = fio_swap64(t->sector);
	350	t->bytes = fio_swap32(t->bytes);
	351	t->action = fio_swap32(t->action);
	352	t->pid = fio_swap32(t->pid);
	353	t->device = fio_swap32(t->device);
	354	t->cpu = fio_swap32(t->cpu);
	355	t->error = fio_swap16(t->error);
	356	t->pdu_len = fio_swap16(t->pdu_len);
	357	}
	358
8c1fdf04 JA	359	/*
	360	* Load a blktrace file by reading all the blk_io_trace entries, and storing
	361	* them as io_pieces like the fio text version would do.
	362	*/
d95b34a6	363	int load_blktrace(struct thread_data td, const char filename, int need_swap)
fb7b71a3	364	{
a61eddec	365	unsigned long long ttime, delay;
fb7b71a3	366	struct blk_io_trace t;
4241ea8f	367	unsigned long ios[2], skipped_writes;
a61eddec	368	unsigned int cpu;
d84f8d49	369	unsigned int rw_bs[2];
e2887563	370	struct fifo *fifo;
89ac1d48 SL	371	int fd, i;
89ac1d48 SL	372	struct fio_file *f;
fb7b71a3 JA	373
	374	fd = open(filename, O_RDONLY);
	375	if (fd < 0) {
	376	td_verror(td, errno, "open blktrace file");
	377	return 1;
	378	}
	379
e2887563 JA	380	fifo = fifo_alloc(TRACE_FIFO_SIZE);
e2887563 JA	381
6df8adaa JA	382	td->o.size = 0;
6df8adaa JA	383
a61eddec	384	cpu = 0;
d84f8d49 JA	385	ttime = 0;
	386	ios[0] = ios[1] = 0;
	387	rw_bs[0] = rw_bs[1] = 0;
4241ea8f	388	skipped_writes = 0;
fb7b71a3	389	do {
e2887563	390	int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
fb7b71a3	391
e2887563	392	if (ret < 0)
8c1fdf04	393	goto err;
e2887563 JA	394	else if (!ret)
	395	break;
	396	else if (ret < (int) sizeof(t)) {
	397	log_err("fio: short fifo get\n");
fb7b71a3	398	break;
fb7b71a3 JA	399	}
fb7b71a3 JA	400
d95b34a6 JA	401	if (need_swap)
	402	byteswap_trace(&t);
	403
fb7b71a3	404	if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
5ec10eaa JA	405	log_err("fio: bad magic in blktrace data: %x\n",
5ec10eaa JA	406	t.magic);
8c1fdf04	407	goto err;
fb7b71a3 JA	408	}
fb7b71a3 JA	409	if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
5ec10eaa JA	410	log_err("fio: bad blktrace version %d\n",
5ec10eaa JA	411	t.magic & 0xff);
8c1fdf04	412	goto err;
fb7b71a3	413	}
f12b323f JA	414	ret = discard_pdu(td, fifo, fd, &t);
f12b323f JA	415	if (ret < 0) {
fb7b71a3	416	td_verror(td, ret, "blktrace lseek");
8c1fdf04	417	goto err;
f12b323f JA	418	} else if (t.pdu_len != ret) {
	419	log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
	420	goto err;
fb7b71a3	421	}
691c8fb0 JA	422	if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
	423	if (!ttime) {
	424	ttime = t.time;
	425	cpu = t.cpu;
	426	}
	427
	428	delay = 0;
	429	if (cpu == t.cpu)
	430	delay = t.time - ttime;
	431	if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
	432	skipped_writes++;
64bbb865 DN	433	else {
	434	/*
	435	* set delay to zero if no_stall enabled for
	436	* fast replay
	437	*/
	438	if (td->o.no_stall)
	439	delay = 0;
	440
691c8fb0	441	handle_trace(td, &t, delay, ios, rw_bs);
64bbb865	442	}
691c8fb0	443
8c1fdf04	444	ttime = t.time;
a61eddec	445	cpu = t.cpu;
a6edd638 JA	446	} else {
a6edd638 JA	447	delay = 0;
4241ea8f	448	handle_trace(td, &t, delay, ios, rw_bs);
a6edd638	449	}
fb7b71a3 JA	450	} while (1);
fb7b71a3 JA	451
85a47ca2 JA	452	for (i = 0; i < td->files_index; i++) {
85a47ca2 JA	453	f= td->files[i];
89ac1d48	454	trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
85a47ca2	455	}
89ac1d48	456
38470f85	457	fifo_free(fifo);
fb7b71a3	458	close(fd);
8c1fdf04	459
4241ea8f	460	if (skipped_writes)
5ec10eaa JA	461	log_err("fio: %s skips replay of %lu writes due to read-only\n",
5ec10eaa JA	462	td->o.name, skipped_writes);
4241ea8f	463
8c1fdf04 JA	464	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	465	log_err("fio: found no ios in blktrace data\n");
	466	return 1;
d84f8d49	467	} else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
8c1fdf04	468	td->o.td_ddir = TD_DDIR_READ;
d84f8d49 JA	469	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
d84f8d49 JA	470	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
8c1fdf04	471	td->o.td_ddir = TD_DDIR_WRITE;
d84f8d49 JA	472	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
d84f8d49 JA	473	} else {
8c1fdf04	474	td->o.td_ddir = TD_DDIR_RW;
d84f8d49 JA	475	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	476	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	477	}
8c1fdf04 JA	478
	479	/*
	480	* We need to do direct/raw ios to the device, to avoid getting
	481	* read-ahead in our way.
	482	*/
	483	td->o.odirect = 1;
	484
fb7b71a3	485	return 0;
8c1fdf04 JA	486	err:
8c1fdf04 JA	487	close(fd);
38470f85	488	fifo_free(fifo);
8c1fdf04	489	return 1;
fb7b71a3	490	}