[fio.git] / blktrace.c

/*
 * blktrace support code for fio
 */
#include <stdio.h>
#include <stdlib.h>
#include <sys/stat.h>
#include <sys/ioctl.h>
#include <linux/fs.h>
#include <dirent.h>

#include "flist.h"
#include "fio.h"
#include "blktrace.h"
#include "blktrace_api.h"
#include "oslib/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 get_dev_blocksize(const char *dev, unsigned int *bs)
{
	int fd;

	fd = open(dev, O_RDONLY);
	if (fd < 0)
		return 1;

	if (ioctl(fd, BLKSSZGET, bs) < 0) {
		close(fd);
		return 1;
	}

	close(fd);
	return 0;
}

static int trace_add_file(struct thread_data *td, __u32 device,
			  unsigned int *bs)
{
	static unsigned int last_maj, last_min, last_fileno, last_bs;
	unsigned int maj = FMAJOR(device);
	unsigned int min = FMINOR(device);
	struct fio_file *f;
	unsigned int i;
	char dev[256];

	if (last_maj == maj && last_min == min) {
		*bs = last_bs;
		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;
			last_bs = f->bs;
			goto out;
		}
	}

	strcpy(dev, "/dev");
	if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
		unsigned int this_bs;
		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);

		if (get_dev_blocksize(dev, &this_bs))
			this_bs = 512;

		td->o.open_files++;
		td->files[fileno]->major = maj;
		td->files[fileno]->minor = min;
		td->files[fileno]->bs = this_bs;
		trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);

		last_fileno = fileno;
		last_bs = this_bs;
	}

out:
	*bs = last_bs;
	return last_fileno;
}

static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
{
	if (!o->replay_align)
		return;

	t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
}

/*
 * 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, unsigned int bs)
{
	struct io_piece *ipo = malloc(sizeof(*ipo));

	init_ipo(ipo);

	ipo->offset = offset * bs;
	if (td->o.replay_scale)
		ipo->offset = ipo->offset / td->o.replay_scale;
	ipo_bytes_align(td->o.replay_align, ipo);
	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 *rw_bs)
{
	struct io_piece *ipo = malloc(sizeof(*ipo));
	unsigned int bs;
	int fileno;

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

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

	td->o.size += t->bytes;

	memset(ipo, 0, sizeof(*ipo));
	INIT_FLIST_HEAD(&ipo->list);

	ipo->offset = t->sector * bs;
	if (td->o.replay_scale)
		ipo->offset = ipo->offset / td->o.replay_scale;
	ipo_bytes_align(td->o.replay_align, ipo);
	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 *rw_bs)
{
	unsigned int bs;
	int rw;
	int fileno;

	fileno = trace_add_file(td, t->device, &bs);

	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;

	if (t->bytes > rw_bs[rw])
		rw_bs[rw] = t->bytes;

	ios[rw]++;
	td->o.size += t->bytes;
	store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
}

/*
 * 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 = 0;

	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;
		}
	}

	t_bytes_align(&td->o, t);

	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_each_file(td, f, 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 <sys/ioctl.h>
	8	#include <linux/fs.h>
	9	#include <dirent.h>
	10
	11	#include "flist.h"
	12	#include "fio.h"
	13	#include "blktrace.h"
	14	#include "blktrace_api.h"
	15	#include "oslib/linux-dev-lookup.h"
	16
	17	#define TRACE_FIFO_SIZE 8192
	18
	19	/*
	20	* fifo refill frontend, to avoid reading data in trace sized bites
	21	*/
	22	static int refill_fifo(struct thread_data td, struct fifo fifo, int fd)
	23	{
	24	char buf[TRACE_FIFO_SIZE];
	25	unsigned int total;
	26	int ret;
	27
	28	total = sizeof(buf);
	29	if (total > fifo_room(fifo))
	30	total = fifo_room(fifo);
	31
	32	ret = read(fd, buf, total);
	33	if (ret < 0) {
	34	td_verror(td, errno, "read blktrace file");
	35	return -1;
	36	}
	37
	38	if (ret > 0)
	39	ret = fifo_put(fifo, buf, ret);
	40
	41	dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
	42	return ret;
	43	}
	44
	45	/*
	46	* Retrieve 'len' bytes from the fifo, refilling if necessary.
	47	*/
	48	static int trace_fifo_get(struct thread_data td, struct fifo fifo, int fd,
	49	void *buf, unsigned int len)
	50	{
	51	if (fifo_len(fifo) < len) {
	52	int ret = refill_fifo(td, fifo, fd);
	53
	54	if (ret < 0)
	55	return ret;
	56	}
	57
	58	return fifo_get(fifo, buf, len);
	59	}
	60
	61	/*
	62	* Just discard the pdu by seeking past it.
	63	*/
	64	static int discard_pdu(struct thread_data td, struct fifo fifo, int fd,
	65	struct blk_io_trace *t)
	66	{
	67	if (t->pdu_len == 0)
	68	return 0;
	69
	70	dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
	71	return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
	72	}
	73
	74	/*
	75	* Check if this is a blktrace binary data file. We read a single trace
	76	* into memory and check for the magic signature.
	77	*/
	78	int is_blktrace(const char filename, int need_swap)
	79	{
	80	struct blk_io_trace t;
	81	int fd, ret;
	82
	83	fd = open(filename, O_RDONLY);
	84	if (fd < 0)
	85	return 0;
	86
	87	ret = read(fd, &t, sizeof(t));
	88	close(fd);
	89
	90	if (ret < 0) {
	91	perror("read blktrace");
	92	return 0;
	93	} else if (ret != sizeof(t)) {
	94	log_err("fio: short read on blktrace file\n");
	95	return 0;
	96	}
	97
	98	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
	99	*need_swap = 0;
	100	return 1;
	101	}
	102
	103	/*
	104	* Maybe it needs to be endian swapped...
	105	*/
	106	t.magic = fio_swap32(t.magic);
	107	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
	108	*need_swap = 1;
	109	return 1;
	110	}
	111
	112	return 0;
	113	}
	114
	115	#define FMINORBITS 20
	116	#define FMINORMASK ((1U << FMINORBITS) - 1)
	117	#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
	118	#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
	119
	120	static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
	121	{
	122	struct io_piece *ipo;
	123
	124	ipo = calloc(1, sizeof(*ipo));
	125	init_ipo(ipo);
	126
	127	ipo->ddir = DDIR_INVAL;
	128	ipo->fileno = fileno;
	129	ipo->file_action = action;
	130	flist_add_tail(&ipo->list, &td->io_log_list);
	131	}
	132
	133	static int get_dev_blocksize(const char dev, unsigned int bs)
	134	{
	135	int fd;
	136
	137	fd = open(dev, O_RDONLY);
	138	if (fd < 0)
	139	return 1;
	140
	141	if (ioctl(fd, BLKSSZGET, bs) < 0) {
	142	close(fd);
	143	return 1;
	144	}
	145
	146	close(fd);
	147	return 0;
	148	}
	149
	150	static int trace_add_file(struct thread_data *td, __u32 device,
	151	unsigned int *bs)
	152	{
	153	static unsigned int last_maj, last_min, last_fileno, last_bs;
	154	unsigned int maj = FMAJOR(device);
	155	unsigned int min = FMINOR(device);
	156	struct fio_file *f;
	157	unsigned int i;
	158	char dev[256];
	159
	160	if (last_maj == maj && last_min == min) {
	161	*bs = last_bs;
	162	return last_fileno;
	163	}
	164
	165	last_maj = maj;
	166	last_min = min;
	167
	168	/*
	169	* check for this file in our list
	170	*/
	171	for_each_file(td, f, i) {
	172	if (f->major == maj && f->minor == min) {
	173	last_fileno = f->fileno;
	174	last_bs = f->bs;
	175	goto out;
	176	}
	177	}
	178
	179	strcpy(dev, "/dev");
	180	if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
	181	unsigned int this_bs;
	182	int fileno;
	183
	184	if (td->o.replay_redirect)
	185	dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
	186	" with: %s\n", maj, min,
	187	td->o.replay_redirect);
	188	else
	189	dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
	190
	191	dprint(FD_BLKTRACE, "add devices %s\n", dev);
	192	fileno = add_file_exclusive(td, dev);
	193
	194	if (get_dev_blocksize(dev, &this_bs))
	195	this_bs = 512;
	196
	197	td->o.open_files++;
	198	td->files[fileno]->major = maj;
	199	td->files[fileno]->minor = min;
	200	td->files[fileno]->bs = this_bs;
	201	trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
	202
	203	last_fileno = fileno;
	204	last_bs = this_bs;
	205	}
	206
	207	out:
	208	*bs = last_bs;
	209	return last_fileno;
	210	}
	211
	212	static void t_bytes_align(struct thread_options o, struct blk_io_trace t)
	213	{
	214	if (!o->replay_align)
	215	return;
	216
	217	t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
	218	}
	219
	220	/*
	221	* Store blk_io_trace data in an ipo for later retrieval.
	222	*/
	223	static void store_ipo(struct thread_data *td, unsigned long long offset,
	224	unsigned int bytes, int rw, unsigned long long ttime,
	225	int fileno, unsigned int bs)
	226	{
	227	struct io_piece ipo = malloc(sizeof(ipo));
	228
	229	init_ipo(ipo);
	230
	231	ipo->offset = offset * bs;
	232	if (td->o.replay_scale)
	233	ipo->offset = ipo->offset / td->o.replay_scale;
	234	ipo_bytes_align(td->o.replay_align, ipo);
	235	ipo->len = bytes;
	236	ipo->delay = ttime / 1000;
	237	if (rw)
	238	ipo->ddir = DDIR_WRITE;
	239	else
	240	ipo->ddir = DDIR_READ;
	241	ipo->fileno = fileno;
	242
	243	dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
	244	ipo->ddir, ipo->offset,
	245	ipo->len, ipo->delay);
	246	queue_io_piece(td, ipo);
	247	}
	248
	249	static void handle_trace_notify(struct blk_io_trace *t)
	250	{
	251	switch (t->action) {
	252	case BLK_TN_PROCESS:
	253	dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
	254	t->action, t->pid);
	255	break;
	256	case BLK_TN_TIMESTAMP:
	257	dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
	258	t->action, t->pid);
	259	break;
	260	case BLK_TN_MESSAGE:
	261	break;
	262	default:
	263	dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
	264	break;
	265	}
	266	}
	267
	268	static void handle_trace_discard(struct thread_data *td,
	269	struct blk_io_trace *t,
	270	unsigned long long ttime,
	271	unsigned long ios, unsigned int rw_bs)
	272	{
	273	struct io_piece ipo = malloc(sizeof(ipo));
	274	unsigned int bs;
	275	int fileno;
	276
	277	init_ipo(ipo);
	278	fileno = trace_add_file(td, t->device, &bs);
	279
	280	ios[DDIR_TRIM]++;
	281	if (t->bytes > rw_bs[DDIR_TRIM])
	282	rw_bs[DDIR_TRIM] = t->bytes;
	283
	284	td->o.size += t->bytes;
	285
	286	memset(ipo, 0, sizeof(*ipo));
	287	INIT_FLIST_HEAD(&ipo->list);
	288
	289	ipo->offset = t->sector * bs;
	290	if (td->o.replay_scale)
	291	ipo->offset = ipo->offset / td->o.replay_scale;
	292	ipo_bytes_align(td->o.replay_align, ipo);
	293	ipo->len = t->bytes;
	294	ipo->delay = ttime / 1000;
	295	ipo->ddir = DDIR_TRIM;
	296	ipo->fileno = fileno;
	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
	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 *rw_bs)
	307	{
	308	unsigned int bs;
	309	int rw;
	310	int fileno;
	311
	312	fileno = trace_add_file(td, t->device, &bs);
	313
	314	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
	315
	316	if (t->bytes > rw_bs[rw])
	317	rw_bs[rw] = t->bytes;
	318
	319	ios[rw]++;
	320	td->o.size += t->bytes;
	321	store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
	322	}
	323
	324	/*
	325	* We only care for queue traces, most of the others are side effects
	326	* due to internal workings of the block layer.
	327	*/
	328	static void handle_trace(struct thread_data td, struct blk_io_trace t,
	329	unsigned long ios, unsigned int bs)
	330	{
	331	static unsigned long long last_ttime;
	332	unsigned long long delay = 0;
	333
	334	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
	335	return;
	336
	337	if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
	338	if (!last_ttime \|\| td->o.no_stall) {
	339	last_ttime = t->time;
	340	delay = 0;
	341	} else {
	342	delay = t->time - last_ttime;
	343	last_ttime = t->time;
	344	}
	345	}
	346
	347	t_bytes_align(&td->o, t);
	348
	349	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
	350	handle_trace_notify(t);
	351	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
	352	handle_trace_discard(td, t, delay, ios, bs);
	353	else
	354	handle_trace_fs(td, t, delay, ios, bs);
	355	}
	356
	357	static void byteswap_trace(struct blk_io_trace *t)
	358	{
	359	t->magic = fio_swap32(t->magic);
	360	t->sequence = fio_swap32(t->sequence);
	361	t->time = fio_swap64(t->time);
	362	t->sector = fio_swap64(t->sector);
	363	t->bytes = fio_swap32(t->bytes);
	364	t->action = fio_swap32(t->action);
	365	t->pid = fio_swap32(t->pid);
	366	t->device = fio_swap32(t->device);
	367	t->cpu = fio_swap32(t->cpu);
	368	t->error = fio_swap16(t->error);
	369	t->pdu_len = fio_swap16(t->pdu_len);
	370	}
	371
	372	static int t_is_write(struct blk_io_trace *t)
	373	{
	374	return (t->action & BLK_TC_ACT(BLK_TC_WRITE \| BLK_TC_DISCARD)) != 0;
	375	}
	376
	377	static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
	378	{
	379	if (t->action & BLK_TC_ACT(BLK_TC_READ))
	380	return DDIR_READ;
	381	else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
	382	return DDIR_WRITE;
	383	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
	384	return DDIR_TRIM;
	385
	386	return DDIR_INVAL;
	387	}
	388
	389	static void depth_inc(struct blk_io_trace t, int depth)
	390	{
	391	enum fio_ddir ddir;
	392
	393	ddir = t_get_ddir(t);
	394	if (ddir != DDIR_INVAL)
	395	depth[ddir]++;
	396	}
	397
	398	static void depth_dec(struct blk_io_trace t, int depth)
	399	{
	400	enum fio_ddir ddir;
	401
	402	ddir = t_get_ddir(t);
	403	if (ddir != DDIR_INVAL)
	404	depth[ddir]--;
	405	}
	406
	407	static void depth_end(struct blk_io_trace t, int this_depth, int *depth)
	408	{
	409	enum fio_ddir ddir = DDIR_INVAL;
	410
	411	ddir = t_get_ddir(t);
	412	if (ddir != DDIR_INVAL) {
	413	depth[ddir] = max(depth[ddir], this_depth[ddir]);
	414	this_depth[ddir] = 0;
	415	}
	416	}
	417
	418	/*
	419	* Load a blktrace file by reading all the blk_io_trace entries, and storing
	420	* them as io_pieces like the fio text version would do.
	421	*/
	422	int load_blktrace(struct thread_data td, const char filename, int need_swap)
	423	{
	424	struct blk_io_trace t;
	425	unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
	426	unsigned int rw_bs[DDIR_RWDIR_CNT];
	427	struct fifo *fifo;
	428	int fd, i, old_state;
	429	struct fio_file *f;
	430	int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
	431
	432	fd = open(filename, O_RDONLY);
	433	if (fd < 0) {
	434	td_verror(td, errno, "open blktrace file");
	435	return 1;
	436	}
	437
	438	fifo = fifo_alloc(TRACE_FIFO_SIZE);
	439
	440	old_state = td_bump_runstate(td, TD_SETTING_UP);
	441
	442	td->o.size = 0;
	443
	444	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
	445	ios[i] = 0;
	446	rw_bs[i] = 0;
	447	this_depth[i] = 0;
	448	depth[i] = 0;
	449	}
	450
	451	skipped_writes = 0;
	452	do {
	453	int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
	454
	455	if (ret < 0)
	456	goto err;
	457	else if (!ret)
	458	break;
	459	else if (ret < (int) sizeof(t)) {
	460	log_err("fio: short fifo get\n");
	461	break;
	462	}
	463
	464	if (need_swap)
	465	byteswap_trace(&t);
	466
	467	if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
	468	log_err("fio: bad magic in blktrace data: %x\n",
	469	t.magic);
	470	goto err;
	471	}
	472	if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
	473	log_err("fio: bad blktrace version %d\n",
	474	t.magic & 0xff);
	475	goto err;
	476	}
	477	ret = discard_pdu(td, fifo, fd, &t);
	478	if (ret < 0) {
	479	td_verror(td, ret, "blktrace lseek");
	480	goto err;
	481	} else if (t.pdu_len != ret) {
	482	log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
	483	goto err;
	484	}
	485	if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
	486	if ((t.action & 0xffff) == __BLK_TA_QUEUE)
	487	depth_inc(&t, this_depth);
	488	else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) \|\|
	489	((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
	490	depth_dec(&t, this_depth);
	491	else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
	492	depth_end(&t, this_depth, depth);
	493
	494	if (t_is_write(&t) && read_only) {
	495	skipped_writes++;
	496	continue;
	497	}
	498	}
	499
	500	handle_trace(td, &t, ios, rw_bs);
	501	} while (1);
	502
	503	for_each_file(td, f, i)
	504	trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
	505
	506	fifo_free(fifo);
	507	close(fd);
	508
	509	td_restore_runstate(td, old_state);
	510
	511	if (!td->files_index) {
	512	log_err("fio: did not find replay device(s)\n");
	513	return 1;
	514	}
	515
	516	/*
	517	* For stacked devices, we don't always get a COMPLETE event so
	518	* the depth grows to insane values. Limit it to something sane(r).
	519	*/
	520	max_depth = 0;
	521	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
	522	if (depth[i] > 1024)
	523	depth[i] = 1024;
	524	else if (!depth[i] && ios[i])
	525	depth[i] = 1;
	526	max_depth = max(depth[i], max_depth);
	527	}
	528
	529	if (skipped_writes)
	530	log_err("fio: %s skips replay of %lu writes due to read-only\n",
	531	td->o.name, skipped_writes);
	532
	533	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	534	log_err("fio: found no ios in blktrace data\n");
	535	return 1;
	536	} else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	537	td->o.td_ddir = TD_DDIR_READ;
	538	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	539	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
	540	td->o.td_ddir = TD_DDIR_WRITE;
	541	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	542	} else {
	543	td->o.td_ddir = TD_DDIR_RW;
	544	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
	545	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
	546	td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
	547	}
	548
	549	/*
	550	* We need to do direct/raw ios to the device, to avoid getting
	551	* read-ahead in our way. But only do so if the minimum block size
	552	* is a multiple of 4k, otherwise we don't know if it's safe to do so.
	553	*/
	554	if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
	555	td->o.odirect = 1;
	556
	557	/*
	558	* If depth wasn't manually set, use probed depth
	559	*/
	560	if (!fio_option_is_set(&td->o, iodepth))
	561	td->o.iodepth = td->o.iodepth_low = max_depth;
	562
	563	return 0;
	564	err:
	565	close(fd);
	566	fifo_free(fifo);
	567	return 1;
	568	}