[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"
#include "lib/linux-dev-lookup.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;
}

#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 (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
		int fileno;

		if (td->o.replay_redirect)
			dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
					" with: %s\n", maj, min,
					td->o.replay_redirect);
		else
			dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);

		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;
}

static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
{
	if (t->action & BLK_TC_ACT(BLK_TC_READ))
		return DDIR_READ;
	else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
		return DDIR_WRITE;
	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
		return DDIR_TRIM;

	return DDIR_INVAL;
}

static void depth_inc(struct blk_io_trace *t, int *depth)
{
	enum fio_ddir ddir;

	ddir = t_get_ddir(t);
	if (ddir != DDIR_INVAL)
		depth[ddir]++;
}

static void depth_dec(struct blk_io_trace *t, int *depth)
{
	enum fio_ddir ddir;

	ddir = t_get_ddir(t);
	if (ddir != DDIR_INVAL)
		depth[ddir]--;
}

static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
{
	enum fio_ddir ddir = DDIR_INVAL;

	ddir = t_get_ddir(t);
	if (ddir != DDIR_INVAL) {
		depth[ddir] = max(depth[ddir], this_depth[ddir]);
		this_depth[ddir] = 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[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_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;

	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
		ios[i] = 0;
		rw_bs[i] = 0;
		this_depth[i] = 0;
		depth[i] = 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 ((t.action & 0xffff) == __BLK_TA_QUEUE)
				depth_inc(&t, this_depth);
			else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
				((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
				depth_dec(&t, this_depth);
			else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
				depth_end(&t, this_depth, depth);

			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).
	 */
	max_depth = 0;
	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
		if (depth[i] > 1024)
			depth[i] = 1024;
		else if (!depth[i] && ios[i])
			depth[i] = 1;
		max_depth = max(depth[i], max_depth);
	}

	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. But only do so if the minimum block size
	 * is a multiple of 4k, otherwise we don't know if it's safe to do so.
	 */
	if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
		td->o.odirect = 1;

	/*
	 * If depth wasn't manually set, use probed depth
	 */
	if (!fio_option_is_set(&td->o, iodepth))
		td->o.iodepth = td->o.iodepth_low = max_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	#include "lib/linux-dev-lookup.h"
	13
	14	#define TRACE_FIFO_SIZE 8192
	15
	16	/*
	17	* fifo refill frontend, to avoid reading data in trace sized bites
	18	*/
	19	static int refill_fifo(struct thread_data td, struct fifo fifo, int fd)
	20	{
	21	char buf[TRACE_FIFO_SIZE];
	22	unsigned int total;
	23	int ret;
	24
	25	total = sizeof(buf);
	26	if (total > fifo_room(fifo))
	27	total = fifo_room(fifo);
	28
	29	ret = read(fd, buf, total);
	30	if (ret < 0) {
	31	td_verror(td, errno, "read blktrace file");
	32	return -1;
	33	}
	34
	35	if (ret > 0)
	36	ret = fifo_put(fifo, buf, ret);
	37
	38	dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
	39	return ret;
	40	}
	41
	42	/*
	43	* Retrieve 'len' bytes from the fifo, refilling if necessary.
	44	*/
	45	static int trace_fifo_get(struct thread_data td, struct fifo fifo, int fd,
	46	void *buf, unsigned int len)
	47	{
	48	if (fifo_len(fifo) < len) {
	49	int ret = refill_fifo(td, fifo, fd);
	50
	51	if (ret < 0)
	52	return ret;
	53	}
	54
	55	return fifo_get(fifo, buf, len);
	56	}
	57
	58	/*
	59	* Just discard the pdu by seeking past it.
	60	*/
	61	static int discard_pdu(struct thread_data td, struct fifo fifo, int fd,
	62	struct blk_io_trace *t)
	63	{
	64	if (t->pdu_len == 0)
	65	return 0;
	66
	67	dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
	68	return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
	69	}
	70
	71	/*
	72	* Check if this is a blktrace binary data file. We read a single trace
	73	* into memory and check for the magic signature.
	74	*/
	75	int is_blktrace(const char filename, int need_swap)
	76	{
	77	struct blk_io_trace t;
	78	int fd, ret;
	79
	80	fd = open(filename, O_RDONLY);
	81	if (fd < 0)
	82	return 0;
	83
	84	ret = read(fd, &t, sizeof(t));
	85	close(fd);
	86
	87	if (ret < 0) {
	88	perror("read blktrace");
	89	return 0;
	90	} else if (ret != sizeof(t)) {
	91	log_err("fio: short read on blktrace file\n");
	92	return 0;
	93	}
	94
	95	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
	96	*need_swap = 0;
	97	return 1;
	98	}
	99
	100	/*
	101	* Maybe it needs to be endian swapped...
	102	*/
	103	t.magic = fio_swap32(t.magic);
	104	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
	105	*need_swap = 1;
	106	return 1;
	107	}
	108
	109	return 0;
	110	}
	111
	112	#define FMINORBITS 20
	113	#define FMINORMASK ((1U << FMINORBITS) - 1)
	114	#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
	115	#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
	116
	117	static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
	118	{
	119	struct io_piece *ipo;
	120
	121	ipo = calloc(1, sizeof(*ipo));
	122	init_ipo(ipo);
	123
	124	ipo->ddir = DDIR_INVAL;
	125	ipo->fileno = fileno;
	126	ipo->file_action = action;
	127	flist_add_tail(&ipo->list, &td->io_log_list);
	128	}
	129
	130	static int trace_add_file(struct thread_data *td, __u32 device)
	131	{
	132	static unsigned int last_maj, last_min, last_fileno;
	133	unsigned int maj = FMAJOR(device);
	134	unsigned int min = FMINOR(device);
	135	struct fio_file *f;
	136	char dev[256];
	137	unsigned int i;
	138
	139	if (last_maj == maj && last_min == min)
	140	return last_fileno;
	141
	142	last_maj = maj;
	143	last_min = min;
	144
	145	/*
	146	* check for this file in our list
	147	*/
	148	for_each_file(td, f, i)
	149	if (f->major == maj && f->minor == min) {
	150	last_fileno = f->fileno;
	151	return last_fileno;
	152	}
	153
	154	strcpy(dev, "/dev");
	155	if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
	156	int fileno;
	157
	158	if (td->o.replay_redirect)
	159	dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
	160	" with: %s\n", maj, min,
	161	td->o.replay_redirect);
	162	else
	163	dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
	164
	165	dprint(FD_BLKTRACE, "add devices %s\n", dev);
	166	fileno = add_file_exclusive(td, dev);
	167	td->o.open_files++;
	168	td->files[fileno]->major = maj;
	169	td->files[fileno]->minor = min;
	170	trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
	171	last_fileno = fileno;
	172	}
	173
	174	return last_fileno;
	175	}
	176
	177	/*
	178	* Store blk_io_trace data in an ipo for later retrieval.
	179	*/
	180	static void store_ipo(struct thread_data *td, unsigned long long offset,
	181	unsigned int bytes, int rw, unsigned long long ttime,
	182	int fileno)
	183	{
	184	struct io_piece ipo = malloc(sizeof(ipo));
	185
	186	init_ipo(ipo);
	187
	188	/*
	189	* the 512 is wrong here, it should be the hardware sector size...
	190	*/
	191	ipo->offset = offset * 512;
	192	ipo->len = bytes;
	193	ipo->delay = ttime / 1000;
	194	if (rw)
	195	ipo->ddir = DDIR_WRITE;
	196	else
	197	ipo->ddir = DDIR_READ;
	198	ipo->fileno = fileno;
	199
	200	dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
	201	ipo->ddir, ipo->offset,
	202	ipo->len, ipo->delay);
	203	queue_io_piece(td, ipo);
	204	}
	205
	206	static void handle_trace_notify(struct blk_io_trace *t)
	207	{
	208	switch (t->action) {
	209	case BLK_TN_PROCESS:
	210	dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
	211	t->action, t->pid);
	212	break;
	213	case BLK_TN_TIMESTAMP:
	214	dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
	215	t->action, t->pid);
	216	break;
	217	case BLK_TN_MESSAGE:
	218	break;
	219	default:
	220	dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
	221	break;
	222	}
	223	}
	224
	225	static void handle_trace_discard(struct thread_data *td,
	226	struct blk_io_trace *t,
	227	unsigned long long ttime,
	228	unsigned long ios, unsigned int bs)
	229	{
	230	struct io_piece ipo = malloc(sizeof(ipo));
	231	int fileno;
	232
	233	init_ipo(ipo);
	234	fileno = trace_add_file(td, t->device);
	235
	236	ios[DDIR_TRIM]++;
	237	if (t->bytes > bs[DDIR_TRIM])
	238	bs[DDIR_TRIM] = t->bytes;
	239
	240	td->o.size += t->bytes;
	241
	242	memset(ipo, 0, sizeof(*ipo));
	243	INIT_FLIST_HEAD(&ipo->list);
	244
	245	/*
	246	* the 512 is wrong here, it should be the hardware sector size...
	247	*/
	248	ipo->offset = t->sector * 512;
	249	ipo->len = t->bytes;
	250	ipo->delay = ttime / 1000;
	251	ipo->ddir = DDIR_TRIM;
	252	ipo->fileno = fileno;
	253
	254	dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
	255	ipo->offset, ipo->len,
	256	ipo->delay);
	257	queue_io_piece(td, ipo);
	258	}
	259
	260	static void handle_trace_fs(struct thread_data td, struct blk_io_trace t,
	261	unsigned long long ttime, unsigned long *ios,
	262	unsigned int *bs)
	263	{
	264	int rw;
	265	int fileno;
	266
	267	fileno = trace_add_file(td, t->device);
	268
	269	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
	270
	271	if (t->bytes > bs[rw])
	272	bs[rw] = t->bytes;
	273
	274	ios[rw]++;
	275	td->o.size += t->bytes;
	276	store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
	277	}
	278
	279	/*
	280	* We only care for queue traces, most of the others are side effects
	281	* due to internal workings of the block layer.
	282	*/
	283	static void handle_trace(struct thread_data td, struct blk_io_trace t,
	284	unsigned long ios, unsigned int bs)
	285	{
	286	static unsigned long long last_ttime;
	287	unsigned long long delay;
	288
	289	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
	290	return;
	291
	292	if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
	293	if (!last_ttime \|\| td->o.no_stall) {
	294	last_ttime = t->time;
	295	delay = 0;
	296	} else {
	297	delay = t->time - last_ttime;
	298	last_ttime = t->time;
	299	}
	300	}
	301
	302	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
	303	handle_trace_notify(t);
	304	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
	305	handle_trace_discard(td, t, delay, ios, bs);
	306	else
	307	handle_trace_fs(td, t, delay, ios, bs);
	308	}
	309
	310	static void byteswap_trace(struct blk_io_trace *t)
	311	{
	312	t->magic = fio_swap32(t->magic);
	313	t->sequence = fio_swap32(t->sequence);
	314	t->time = fio_swap64(t->time);
	315	t->sector = fio_swap64(t->sector);
	316	t->bytes = fio_swap32(t->bytes);
	317	t->action = fio_swap32(t->action);
	318	t->pid = fio_swap32(t->pid);
	319	t->device = fio_swap32(t->device);
	320	t->cpu = fio_swap32(t->cpu);
	321	t->error = fio_swap16(t->error);
	322	t->pdu_len = fio_swap16(t->pdu_len);
	323	}
	324
	325	static int t_is_write(struct blk_io_trace *t)
	326	{
	327	return (t->action & BLK_TC_ACT(BLK_TC_WRITE \| BLK_TC_DISCARD)) != 0;
	328	}
	329
	330	static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
	331	{
	332	if (t->action & BLK_TC_ACT(BLK_TC_READ))
	333	return DDIR_READ;
	334	else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
	335	return DDIR_WRITE;
	336	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
	337	return DDIR_TRIM;
	338
	339	return DDIR_INVAL;
	340	}
	341
	342	static void depth_inc(struct blk_io_trace t, int depth)
	343	{
	344	enum fio_ddir ddir;
	345
	346	ddir = t_get_ddir(t);
	347	if (ddir != DDIR_INVAL)
	348	depth[ddir]++;
	349	}
	350
	351	static void depth_dec(struct blk_io_trace t, int depth)
	352	{
	353	enum fio_ddir ddir;
	354
	355	ddir = t_get_ddir(t);
	356	if (ddir != DDIR_INVAL)
	357	depth[ddir]--;
	358	}
	359
	360	static void depth_end(struct blk_io_trace t, int this_depth, int *depth)
	361	{
	362	enum fio_ddir ddir = DDIR_INVAL;
	363
	364	ddir = t_get_ddir(t);
	365	if (ddir != DDIR_INVAL) {
	366	depth[ddir] = max(depth[ddir], this_depth[ddir]);
	367	this_depth[ddir] = 0;
	368	}
	369	}
	370
	371	/*
	372	* Load a blktrace file by reading all the blk_io_trace entries, and storing
	373	* them as io_pieces like the fio text version would do.
	374	*/
	375	int load_blktrace(struct thread_data td, const char filename, int need_swap)
	376	{
	377	struct blk_io_trace t;
	378	unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
	379	unsigned int rw_bs[DDIR_RWDIR_CNT];
	380	struct fifo *fifo;
	381	int fd, i, old_state;
	382	struct fio_file *f;
	383	int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
	384
	385	fd = open(filename, O_RDONLY);
	386	if (fd < 0) {
	387	td_verror(td, errno, "open blktrace file");
	388	return 1;
	389	}
	390
	391	fifo = fifo_alloc(TRACE_FIFO_SIZE);
	392
	393	old_state = td_bump_runstate(td, TD_SETTING_UP);
	394
	395	td->o.size = 0;
	396
	397	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
	398	ios[i] = 0;
	399	rw_bs[i] = 0;
	400	this_depth[i] = 0;
	401	depth[i] = 0;
	402	}
	403
	404	skipped_writes = 0;
	405	do {
	406	int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
	407
	408	if (ret < 0)
	409	goto err;
	410	else if (!ret)
	411	break;
	412	else if (ret < (int) sizeof(t)) {
	413	log_err("fio: short fifo get\n");
	414	break;
	415	}
	416
	417	if (need_swap)
	418	byteswap_trace(&t);
	419
	420	if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
	421	log_err("fio: bad magic in blktrace data: %x\n",
	422	t.magic);
	423	goto err;
	424	}
	425	if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
	426	log_err("fio: bad blktrace version %d\n",
	427	t.magic & 0xff);
	428	goto err;
	429	}
	430	ret = discard_pdu(td, fifo, fd, &t);
	431	if (ret < 0) {
	432	td_verror(td, ret, "blktrace lseek");
	433	goto err;
	434	} else if (t.pdu_len != ret) {
	435	log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
	436	goto err;
	437	}
	438	if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
	439	if ((t.action & 0xffff) == __BLK_TA_QUEUE)
	440	depth_inc(&t, this_depth);
	441	else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) \|\|
	442	((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
	443	depth_dec(&t, this_depth);
	444	else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
	445	depth_end(&t, this_depth, depth);
	446
	447	if (t_is_write(&t) && read_only) {
	448	skipped_writes++;
	449	continue;
	450	}
	451	}
	452
	453	handle_trace(td, &t, ios, rw_bs);
	454	} while (1);
	455
	456	for (i = 0; i < td->files_index; i++) {
	457	f = td->files[i];
	458	trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
	459	}
	460
	461	fifo_free(fifo);
	462	close(fd);
	463
	464	td_restore_runstate(td, old_state);
	465
	466	if (!td->files_index) {
	467	log_err("fio: did not find replay device(s)\n");
	468	return 1;
	469	}
	470
	471	/*
	472	* For stacked devices, we don't always get a COMPLETE event so
	473	* the depth grows to insane values. Limit it to something sane(r).
	474	*/
	475	max_depth = 0;
	476	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
	477	if (depth[i] > 1024)
	478	depth[i] = 1024;
	479	else if (!depth[i] && ios[i])
	480	depth[i] = 1;
	481	max_depth = max(depth[i], max_depth);
	482	}
	483
	484	if (skipped_writes)
	485	log_err("fio: %s skips replay of %lu writes due to read-only\n",
	486	td->o.name, skipped_writes);
	487
	488	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	489	log_err("fio: found no ios in blktrace data\n");
	490	return 1;
	491	} else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	492	td->o.td_ddir = TD_DDIR_READ;
	493	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	494	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
	495	td->o.td_ddir = TD_DDIR_WRITE;
	496	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	497	} else {
	498	td->o.td_ddir = TD_DDIR_RW;
	499	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	500	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	501	td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
	502	}
	503
	504	/*
	505	* We need to do direct/raw ios to the device, to avoid getting
	506	* read-ahead in our way. But only do so if the minimum block size
	507	* is a multiple of 4k, otherwise we don't know if it's safe to do so.
	508	*/
	509	if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
	510	td->o.odirect = 1;
	511
	512	/*
	513	* If depth wasn't manually set, use probed depth
	514	*/
	515	if (!fio_option_is_set(&td->o, iodepth))
	516	td->o.iodepth = td->o.iodepth_low = max_depth;
	517
	518	return 0;
	519	err:
	520	close(fd);
	521	fifo_free(fifo);
	522	return 1;
	523	}