[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\n", 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);
		td->o.open_files++;
		td->files[fileno]->major = maj;
		td->files[fileno]->minor = min;
		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:
		dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
				t->action, t->pid);
		break;
	case BLK_TN_TIMESTAMP:
		dprint(FD_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, unsigned int *bs)
{
	struct io_piece *ipo = malloc(sizeof(*ipo));
	int fileno;

	init_ipo(ipo);
	fileno = trace_add_file(td, t->device);

	ios[DDIR_TRIM]++;
	if (t->bytes > bs[DDIR_TRIM])
		bs[DDIR_TRIM] = t->bytes;

	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 *ios, unsigned int *bs)
{
	static unsigned long long last_ttime;
	unsigned long long delay;

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

	if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
		if (!last_ttime || td->o.no_stall) {
			last_ttime = t->time;
			delay = 0;
		} else {
			delay = t->time - last_ttime;
			last_ttime = t->time;
		}
	}

	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, delay, ios, bs);
	else
		handle_trace_fs(td, t, delay, 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);
}

static int t_is_write(struct blk_io_trace *t)
{
	return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
}

/*
 * 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)
{
	struct blk_io_trace t;
	unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
	unsigned int rw_bs[DDIR_RWDIR_CNT];
	struct fifo *fifo;
	int fd, i, old_state;
	struct fio_file *f;
	int this_depth, depth;

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

	fifo = fifo_alloc(TRACE_FIFO_SIZE);

	old_state = td_bump_runstate(td, TD_SETTING_UP);

	td->o.size = 0;

	ios[0] = ios[1] = 0;
	rw_bs[0] = rw_bs[1] = 0;
	skipped_writes = 0;
	this_depth = depth = 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 ((t.action & 0xffff) == __BLK_TA_QUEUE)
				this_depth++;
			else if ((t.action & 0xffff) == __BLK_TA_COMPLETE) {
				depth = max(depth, this_depth);
				this_depth = 0;
			}

			if (t_is_write(&t) && read_only) {
				skipped_writes++;
				continue;
			}
		}

		handle_trace(td, &t, 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);

	td_restore_runstate(td, old_state);

	if (!td->files_index) {
		log_err("fio: did not find replay device(s)\n");
		return 1;
	}

	/*
	 * For stacked devices, we don't always get a COMPLETE event so
	 * the depth grows to insane values. Limit it to something sane(r).
	 */
	if (!depth || depth > 1024)
		depth = 1024;

	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_WRITE]) {
		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];
		td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
	}

	/*
	 * We need to do direct/raw ios to the device, to avoid getting
	 * read-ahead in our way.
	 */
	td->o.odirect = 1;

	/*
	 * we don't know if this option was set or not. it defaults to 1,
	 * so we'll just guess that we should override it if it's still 1
	 */
	if (td->o.iodepth != 1)
		td->o.iodepth = depth;

	return 0;
err:
	close(fd);
	fifo_free(fifo);
	return 1;
}
Commit	Line	Data
	1	/*
	2	* blktrace support code for fio
	3	*/
	4	#include <stdio.h>
	5	#include <stdlib.h>
	6	#include <sys/stat.h>
	7	#include <dirent.h>
	8
	9	#include "flist.h"
	10	#include "fio.h"
	11	#include "blktrace_api.h"
	12
	13	#define TRACE_FIFO_SIZE 8192
	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];
	21	unsigned int total;
	22	int ret;
	23
	24	total = sizeof(buf);
	25	if (total > fifo_room(fifo))
	26	total = fifo_room(fifo);
	27
	28	ret = read(fd, buf, total);
	29	if (ret < 0) {
	30	td_verror(td, errno, "read blktrace file");
	31	return -1;
	32	}
	33
	34	if (ret > 0)
	35	ret = fifo_put(fifo, buf, ret);
	36
	37	dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
	38	return ret;
	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	{
	47	if (fifo_len(fifo) < len) {
	48	int ret = refill_fifo(td, fifo, fd);
	49
	50	if (ret < 0)
	51	return ret;
	52	}
	53
	54	return fifo_get(fifo, buf, len);
	55	}
	56
	57	/*
	58	* Just discard the pdu by seeking past it.
	59	*/
	60	static int discard_pdu(struct thread_data td, struct fifo fifo, int fd,
	61	struct blk_io_trace *t)
	62	{
	63	if (t->pdu_len == 0)
	64	return 0;
	65
	66	dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
	67	return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
	68	}
	69
	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	*/
	74	int is_blktrace(const char filename, int need_swap)
	75	{
	76	struct blk_io_trace t;
	77	int fd, ret;
	78
	79	fd = open(filename, O_RDONLY);
	80	if (fd < 0)
	81	return 0;
	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	*need_swap = 0;
	96	return 1;
	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	}
	107
	108	return 0;
	109	}
	110
	111	static int lookup_device(struct thread_data td, char path, unsigned int maj,
	112	unsigned int min)
	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
	129	sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name);
	130	if (lstat(full_path, &st) == -1) {
	131	perror("lstat");
	132	break;
	133	}
	134
	135	if (S_ISDIR(st.st_mode)) {
	136	found = lookup_device(td, full_path, maj, min);
	137	if (found) {
	138	strcpy(path, full_path);
	139	break;
	140	}
	141	}
	142
	143	if (!S_ISBLK(st.st_mode))
	144	continue;
	145
	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\n", maj, min,
	154	td->o.replay_redirect);
	155	strcpy(path, td->o.replay_redirect);
	156	found = 1;
	157	break;
	158	}
	159
	160	if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
	161	dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
	162	strcpy(path, full_path);
	163	found = 1;
	164	break;
	165	}
	166	}
	167
	168	closedir(D);
	169	return found;
	170	}
	171
	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))
	176
	177	static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
	178	{
	179	struct io_piece *ipo;
	180
	181	ipo = calloc(1, sizeof(*ipo));
	182	init_ipo(ipo);
	183
	184	ipo->ddir = DDIR_INVAL;
	185	ipo->fileno = fileno;
	186	ipo->file_action = action;
	187	flist_add_tail(&ipo->list, &td->io_log_list);
	188	}
	189
	190	static int trace_add_file(struct thread_data *td, __u32 device)
	191	{
	192	static unsigned int last_maj, last_min, last_fileno;
	193	unsigned int maj = FMAJOR(device);
	194	unsigned int min = FMINOR(device);
	195	struct fio_file *f;
	196	char dev[256];
	197	unsigned int i;
	198
	199	if (last_maj == maj && last_min == min)
	200	return last_fileno;
	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)
	209	if (f->major == maj && f->minor == min) {
	210	last_fileno = f->fileno;
	211	return last_fileno;
	212	}
	213
	214	strcpy(dev, "/dev");
	215	if (lookup_device(td, dev, maj, min)) {
	216	int fileno;
	217
	218	dprint(FD_BLKTRACE, "add devices %s\n", dev);
	219	fileno = add_file_exclusive(td, dev);
	220	td->o.open_files++;
	221	td->files[fileno]->major = maj;
	222	td->files[fileno]->minor = min;
	223	trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
	224	last_fileno = fileno;
	225	}
	226
	227	return last_fileno;
	228	}
	229
	230	/*
	231	* Store blk_io_trace data in an ipo for later retrieval.
	232	*/
	233	static void store_ipo(struct thread_data *td, unsigned long long offset,
	234	unsigned int bytes, int rw, unsigned long long ttime,
	235	int fileno)
	236	{
	237	struct io_piece ipo = malloc(sizeof(ipo));
	238
	239	init_ipo(ipo);
	240
	241	/*
	242	* the 512 is wrong here, it should be the hardware sector size...
	243	*/
	244	ipo->offset = offset * 512;
	245	ipo->len = bytes;
	246	ipo->delay = ttime / 1000;
	247	if (rw)
	248	ipo->ddir = DDIR_WRITE;
	249	else
	250	ipo->ddir = DDIR_READ;
	251	ipo->fileno = fileno;
	252
	253	dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
	254	ipo->ddir, ipo->offset,
	255	ipo->len, ipo->delay);
	256	queue_io_piece(td, ipo);
	257	}
	258
	259	static void handle_trace_notify(struct blk_io_trace *t)
	260	{
	261	switch (t->action) {
	262	case BLK_TN_PROCESS:
	263	dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
	264	t->action, t->pid);
	265	break;
	266	case BLK_TN_TIMESTAMP:
	267	dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
	268	t->action, t->pid);
	269	break;
	270	case BLK_TN_MESSAGE:
	271	break;
	272	default:
	273	dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
	274	break;
	275	}
	276	}
	277
	278	static void handle_trace_discard(struct thread_data *td,
	279	struct blk_io_trace *t,
	280	unsigned long long ttime,
	281	unsigned long ios, unsigned int bs)
	282	{
	283	struct io_piece ipo = malloc(sizeof(ipo));
	284	int fileno;
	285
	286	init_ipo(ipo);
	287	fileno = trace_add_file(td, t->device);
	288
	289	ios[DDIR_TRIM]++;
	290	if (t->bytes > bs[DDIR_TRIM])
	291	bs[DDIR_TRIM] = t->bytes;
	292
	293	td->o.size += t->bytes;
	294
	295	memset(ipo, 0, sizeof(*ipo));
	296	INIT_FLIST_HEAD(&ipo->list);
	297
	298	/*
	299	* the 512 is wrong here, it should be the hardware sector size...
	300	*/
	301	ipo->offset = t->sector * 512;
	302	ipo->len = t->bytes;
	303	ipo->delay = ttime / 1000;
	304	ipo->ddir = DDIR_TRIM;
	305	ipo->fileno = fileno;
	306
	307	dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
	308	ipo->offset, ipo->len,
	309	ipo->delay);
	310	queue_io_piece(td, ipo);
	311	}
	312
	313	static void handle_trace_fs(struct thread_data td, struct blk_io_trace t,
	314	unsigned long long ttime, unsigned long *ios,
	315	unsigned int *bs)
	316	{
	317	int rw;
	318	int fileno;
	319
	320	fileno = trace_add_file(td, t->device);
	321
	322	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
	323
	324	if (t->bytes > bs[rw])
	325	bs[rw] = t->bytes;
	326
	327	ios[rw]++;
	328	td->o.size += t->bytes;
	329	store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
	330	}
	331
	332	/*
	333	* We only care for queue traces, most of the others are side effects
	334	* due to internal workings of the block layer.
	335	*/
	336	static void handle_trace(struct thread_data td, struct blk_io_trace t,
	337	unsigned long ios, unsigned int bs)
	338	{
	339	static unsigned long long last_ttime;
	340	unsigned long long delay;
	341
	342	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
	343	return;
	344
	345	if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
	346	if (!last_ttime \|\| td->o.no_stall) {
	347	last_ttime = t->time;
	348	delay = 0;
	349	} else {
	350	delay = t->time - last_ttime;
	351	last_ttime = t->time;
	352	}
	353	}
	354
	355	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
	356	handle_trace_notify(t);
	357	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
	358	handle_trace_discard(td, t, delay, ios, bs);
	359	else
	360	handle_trace_fs(td, t, delay, ios, bs);
	361	}
	362
	363	static void byteswap_trace(struct blk_io_trace *t)
	364	{
	365	t->magic = fio_swap32(t->magic);
	366	t->sequence = fio_swap32(t->sequence);
	367	t->time = fio_swap64(t->time);
	368	t->sector = fio_swap64(t->sector);
	369	t->bytes = fio_swap32(t->bytes);
	370	t->action = fio_swap32(t->action);
	371	t->pid = fio_swap32(t->pid);
	372	t->device = fio_swap32(t->device);
	373	t->cpu = fio_swap32(t->cpu);
	374	t->error = fio_swap16(t->error);
	375	t->pdu_len = fio_swap16(t->pdu_len);
	376	}
	377
	378	static int t_is_write(struct blk_io_trace *t)
	379	{
	380	return (t->action & BLK_TC_ACT(BLK_TC_WRITE \| BLK_TC_DISCARD)) != 0;
	381	}
	382
	383	/*
	384	* Load a blktrace file by reading all the blk_io_trace entries, and storing
	385	* them as io_pieces like the fio text version would do.
	386	*/
	387	int load_blktrace(struct thread_data td, const char filename, int need_swap)
	388	{
	389	struct blk_io_trace t;
	390	unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
	391	unsigned int rw_bs[DDIR_RWDIR_CNT];
	392	struct fifo *fifo;
	393	int fd, i, old_state;
	394	struct fio_file *f;
	395	int this_depth, depth;
	396
	397	fd = open(filename, O_RDONLY);
	398	if (fd < 0) {
	399	td_verror(td, errno, "open blktrace file");
	400	return 1;
	401	}
	402
	403	fifo = fifo_alloc(TRACE_FIFO_SIZE);
	404
	405	old_state = td_bump_runstate(td, TD_SETTING_UP);
	406
	407	td->o.size = 0;
	408
	409	ios[0] = ios[1] = 0;
	410	rw_bs[0] = rw_bs[1] = 0;
	411	skipped_writes = 0;
	412	this_depth = depth = 0;
	413	do {
	414	int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
	415
	416	if (ret < 0)
	417	goto err;
	418	else if (!ret)
	419	break;
	420	else if (ret < (int) sizeof(t)) {
	421	log_err("fio: short fifo get\n");
	422	break;
	423	}
	424
	425	if (need_swap)
	426	byteswap_trace(&t);
	427
	428	if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
	429	log_err("fio: bad magic in blktrace data: %x\n",
	430	t.magic);
	431	goto err;
	432	}
	433	if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
	434	log_err("fio: bad blktrace version %d\n",
	435	t.magic & 0xff);
	436	goto err;
	437	}
	438	ret = discard_pdu(td, fifo, fd, &t);
	439	if (ret < 0) {
	440	td_verror(td, ret, "blktrace lseek");
	441	goto err;
	442	} else if (t.pdu_len != ret) {
	443	log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
	444	goto err;
	445	}
	446	if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
	447	if ((t.action & 0xffff) == __BLK_TA_QUEUE)
	448	this_depth++;
	449	else if ((t.action & 0xffff) == __BLK_TA_COMPLETE) {
	450	depth = max(depth, this_depth);
	451	this_depth = 0;
	452	}
	453
	454	if (t_is_write(&t) && read_only) {
	455	skipped_writes++;
	456	continue;
	457	}
	458	}
	459
	460	handle_trace(td, &t, ios, rw_bs);
	461	} while (1);
	462
	463	for (i = 0; i < td->files_index; i++) {
	464	f = td->files[i];
	465	trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
	466	}
	467
	468	fifo_free(fifo);
	469	close(fd);
	470
	471	td_restore_runstate(td, old_state);
	472
	473	if (!td->files_index) {
	474	log_err("fio: did not find replay device(s)\n");
	475	return 1;
	476	}
	477
	478	/*
	479	* For stacked devices, we don't always get a COMPLETE event so
	480	* the depth grows to insane values. Limit it to something sane(r).
	481	*/
	482	if (!depth \|\| depth > 1024)
	483	depth = 1024;
	484
	485	if (skipped_writes)
	486	log_err("fio: %s skips replay of %lu writes due to read-only\n",
	487	td->o.name, skipped_writes);
	488
	489	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	490	log_err("fio: found no ios in blktrace data\n");
	491	return 1;
	492	} else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	493	td->o.td_ddir = TD_DDIR_READ;
	494	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	495	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
	496	td->o.td_ddir = TD_DDIR_WRITE;
	497	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	498	} else {
	499	td->o.td_ddir = TD_DDIR_RW;
	500	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	501	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	502	td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
	503	}
	504
	505	/*
	506	* We need to do direct/raw ios to the device, to avoid getting
	507	* read-ahead in our way.
	508	*/
	509	td->o.odirect = 1;
	510
	511	/*
	512	* we don't know if this option was set or not. it defaults to 1,
	513	* so we'll just guess that we should override it if it's still 1
	514	*/
	515	if (td->o.iodepth != 1)
	516	td->o.iodepth = depth;
	517
	518	return 0;
	519	err:
	520	close(fd);
	521	fifo_free(fifo);
	522	return 1;
	523	}