[fio.git] / blktrace.c

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

	ipo = calloc(1, 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;
	unsigned int bs;
	int fileno;

	ipo = calloc(1, sizeof(*ipo));
	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;

	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;

	/*
	 * Need to figure out why 0 byte writes end up here sometimes, for
	 * now just ignore them.
	 */
	if (!t->bytes)
		return;
	else 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);
}

static void handle_trace_flush(struct thread_data *td, struct blk_io_trace *t,
			       unsigned long long ttime)
{
	struct io_piece *ipo;
	unsigned int bs;
	int fileno;

	ipo = calloc(1, sizeof(*ipo));
	init_ipo(ipo);
	fileno = trace_add_file(td, t->device, &bs);

	ipo->delay = ttime / 1000;
	ipo->ddir = DDIR_SYNC;
	ipo->fileno = fileno;

	dprint(FD_BLKTRACE, "store flush delay=%lu\n", ipo->delay);
	queue_io_piece(td, ipo);
}

/*
 * 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)
			delay = 0;
		else if (td->o.replay_time_scale == 100)
			delay = t->time - last_ttime;
		else {
			double tmp = t->time - last_ttime;
			double scale;

			scale = (double) 100.0 / (double) td->o.replay_time_scale;
			tmp *= scale;
			delay = tmp;
		}
		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 if (t->action & BLK_TC_ACT(BLK_TC_FLUSH))
		handle_trace_flush(td, t, delay);
	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
fb7b71a3 JA	1	/*
	2	* blktrace support code for fio
	3	*/
	4	#include <stdio.h>
	5	#include <stdlib.h>
26e616d0 JA	6	#include <sys/ioctl.h>
26e616d0 JA	7	#include <linux/fs.h>
8c1fdf04	8
01743ee1	9	#include "flist.h"
fb7b71a3	10	#include "fio.h"
a3e59412	11	#include "blktrace.h"
fb7b71a3	12	#include "blktrace_api.h"
984f30c9	13	#include "oslib/linux-dev-lookup.h"
fb7b71a3	14
2da7df1d	15	#define TRACE_FIFO_SIZE 8192
e2887563 JA	16
	17	/*
	18	* fifo refill frontend, to avoid reading data in trace sized bites
	19	*/
	20	static int refill_fifo(struct thread_data td, struct fifo fifo, int fd)
	21	{
	22	char buf[TRACE_FIFO_SIZE];
f12b323f	23	unsigned int total;
e2887563 JA	24	int ret;
e2887563 JA	25
f12b323f JA	26	total = sizeof(buf);
	27	if (total > fifo_room(fifo))
	28	total = fifo_room(fifo);
e2887563	29
f12b323f JA	30	ret = read(fd, buf, total);
	31	if (ret < 0) {
	32	td_verror(td, errno, "read blktrace file");
	33	return -1;
e2887563 JA	34	}
e2887563 JA	35
f12b323f JA	36	if (ret > 0)
	37	ret = fifo_put(fifo, buf, ret);
	38
bd6f78b2	39	dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
f12b323f	40	return ret;
e2887563 JA	41	}
	42
	43	/*
	44	* Retrieve 'len' bytes from the fifo, refilling if necessary.
	45	*/
	46	static int trace_fifo_get(struct thread_data td, struct fifo fifo, int fd,
	47	void *buf, unsigned int len)
	48	{
f12b323f JA	49	if (fifo_len(fifo) < len) {
f12b323f JA	50	int ret = refill_fifo(td, fifo, fd);
e2887563	51
f12b323f JA	52	if (ret < 0)
	53	return ret;
	54	}
e2887563 JA	55
	56	return fifo_get(fifo, buf, len);
	57	}
	58
8c1fdf04 JA	59	/*
	60	* Just discard the pdu by seeking past it.
	61	*/
f12b323f JA	62	static int discard_pdu(struct thread_data td, struct fifo fifo, int fd,
f12b323f JA	63	struct blk_io_trace *t)
fb7b71a3 JA	64	{
	65	if (t->pdu_len == 0)
	66	return 0;
	67
bd6f78b2	68	dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
f12b323f	69	return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
fb7b71a3 JA	70	}
fb7b71a3 JA	71
8c1fdf04 JA	72	/*
	73	* Check if this is a blktrace binary data file. We read a single trace
	74	* into memory and check for the magic signature.
	75	*/
d95b34a6	76	int is_blktrace(const char filename, int need_swap)
fb7b71a3 JA	77	{
	78	struct blk_io_trace t;
	79	int fd, ret;
	80
	81	fd = open(filename, O_RDONLY);
4dced407	82	if (fd < 0)
fb7b71a3	83	return 0;
fb7b71a3 JA	84
	85	ret = read(fd, &t, sizeof(t));
	86	close(fd);
	87
	88	if (ret < 0) {
	89	perror("read blktrace");
	90	return 0;
	91	} else if (ret != sizeof(t)) {
	92	log_err("fio: short read on blktrace file\n");
	93	return 0;
	94	}
	95
d95b34a6 JA	96	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
d95b34a6 JA	97	*need_swap = 0;
fb7b71a3	98	return 1;
d95b34a6 JA	99	}
	100
	101	/*
	102	* Maybe it needs to be endian swapped...
	103	*/
	104	t.magic = fio_swap32(t.magic);
	105	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
	106	*need_swap = 1;
	107	return 1;
	108	}
fb7b71a3 JA	109
	110	return 0;
	111	}
	112
c69aa91f JA	113	#define FMINORBITS 20
	114	#define FMINORMASK ((1U << FMINORBITS) - 1)
	115	#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
	116	#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
eeb9c2aa	117
89ac1d48	118	static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
691c8fb0 JA	119	{
	120	struct io_piece *ipo;
	121
	122	ipo = calloc(1, sizeof(*ipo));
0d29de83	123	init_ipo(ipo);
691c8fb0 JA	124
	125	ipo->ddir = DDIR_INVAL;
	126	ipo->fileno = fileno;
89ac1d48	127	ipo->file_action = action;
01743ee1	128	flist_add_tail(&ipo->list, &td->io_log_list);
691c8fb0 JA	129	}
691c8fb0 JA	130
26e616d0	131	static int get_dev_blocksize(const char dev, unsigned int bs)
5e6c2067	132	{
26e616d0 JA	133	int fd;
	134
	135	fd = open(dev, O_RDONLY);
	136	if (fd < 0)
	137	return 1;
	138
	139	if (ioctl(fd, BLKSSZGET, bs) < 0) {
	140	close(fd);
	141	return 1;
	142	}
	143
	144	close(fd);
	145	return 0;
	146	}
	147
	148	static int trace_add_file(struct thread_data *td, __u32 device,
	149	unsigned int *bs)
	150	{
	151	static unsigned int last_maj, last_min, last_fileno, last_bs;
c69aa91f JA	152	unsigned int maj = FMAJOR(device);
c69aa91f JA	153	unsigned int min = FMINOR(device);
5e6c2067	154	struct fio_file *f;
5e6c2067	155	unsigned int i;
26e616d0	156	char dev[256];
5e6c2067	157
26e616d0 JA	158	if (last_maj == maj && last_min == min) {
26e616d0 JA	159	*bs = last_bs;
89ac1d48	160	return last_fileno;
26e616d0	161	}
5e6c2067 JA	162
	163	last_maj = maj;
	164	last_min = min;
	165
	166	/*
	167	* check for this file in our list
	168	*/
26e616d0	169	for_each_file(td, f, i) {
89ac1d48 SL	170	if (f->major == maj && f->minor == min) {
89ac1d48 SL	171	last_fileno = f->fileno;
26e616d0 JA	172	last_bs = f->bs;
26e616d0 JA	173	goto out;
89ac1d48	174	}
26e616d0	175	}
5e6c2067 JA	176
5e6c2067 JA	177	strcpy(dev, "/dev");
40bafa33	178	if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
26e616d0	179	unsigned int this_bs;
691c8fb0 JA	180	int fileno;
691c8fb0 JA	181
40bafa33 JA	182	if (td->o.replay_redirect)
	183	dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
	184	" with: %s\n", maj, min,
	185	td->o.replay_redirect);
	186	else
	187	dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
	188
bd6f78b2	189	dprint(FD_BLKTRACE, "add devices %s\n", dev);
49ffb4a2	190	fileno = add_file_exclusive(td, dev);
26e616d0 JA	191
	192	if (get_dev_blocksize(dev, &this_bs))
	193	this_bs = 512;
	194
b53f2c54	195	td->o.open_files++;
5903e7b7 JA	196	td->files[fileno]->major = maj;
5903e7b7 JA	197	td->files[fileno]->minor = min;
26e616d0	198	td->files[fileno]->bs = this_bs;
89ac1d48	199	trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
26e616d0	200
89ac1d48	201	last_fileno = fileno;
26e616d0	202	last_bs = this_bs;
bd6f78b2	203	}
f01b34ae	204
26e616d0 JA	205	out:
26e616d0 JA	206	*bs = last_bs;
89ac1d48	207	return last_fileno;
5e6c2067 JA	208	}
5e6c2067 JA	209
0c63576e JA	210	static void t_bytes_align(struct thread_options o, struct blk_io_trace t)
	211	{
	212	if (!o->replay_align)
	213	return;
	214
	215	t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
	216	}
	217
8c1fdf04 JA	218	/*
	219	* Store blk_io_trace data in an ipo for later retrieval.
	220	*/
fdefd987	221	static void store_ipo(struct thread_data *td, unsigned long long offset,
89ac1d48	222	unsigned int bytes, int rw, unsigned long long ttime,
26e616d0	223	int fileno, unsigned int bs)
fdefd987	224	{
8812d7f2	225	struct io_piece *ipo;
fdefd987	226
8812d7f2	227	ipo = calloc(1, sizeof(*ipo));
0d29de83 JA	228	init_ipo(ipo);
0d29de83 JA	229
26e616d0	230	ipo->offset = offset * bs;
0c63576e JA	231	if (td->o.replay_scale)
0c63576e JA	232	ipo->offset = ipo->offset / td->o.replay_scale;
a79f17bf	233	ipo_bytes_align(td->o.replay_align, ipo);
fdefd987	234	ipo->len = bytes;
8c1fdf04	235	ipo->delay = ttime / 1000;
fdefd987 JA	236	if (rw)
	237	ipo->ddir = DDIR_WRITE;
	238	else
	239	ipo->ddir = DDIR_READ;
89ac1d48	240	ipo->fileno = fileno;
fdefd987	241
bd6f78b2 JA	242	dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
	243	ipo->ddir, ipo->offset,
	244	ipo->len, ipo->delay);
691c8fb0	245	queue_io_piece(td, ipo);
fdefd987 JA	246	}
fdefd987 JA	247
0b9d69ec	248	static void handle_trace_notify(struct blk_io_trace *t)
cd991b9e	249	{
691c8fb0 JA	250	switch (t->action) {
691c8fb0 JA	251	case BLK_TN_PROCESS:
24653680	252	dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
d95b34a6	253	t->action, t->pid);
691c8fb0 JA	254	break;
691c8fb0 JA	255	case BLK_TN_TIMESTAMP:
24653680	256	dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
d95b34a6	257	t->action, t->pid);
691c8fb0	258	break;
ff58fced JA	259	case BLK_TN_MESSAGE:
ff58fced JA	260	break;
691c8fb0 JA	261	default:
	262	dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
	263	break;
	264	}
	265	}
5b3023b8	266
24653680 JA	267	static void handle_trace_discard(struct thread_data *td,
	268	struct blk_io_trace *t,
	269	unsigned long long ttime,
26e616d0	270	unsigned long ios, unsigned int rw_bs)
ff58fced	271	{
8812d7f2	272	struct io_piece *ipo;
26e616d0	273	unsigned int bs;
89ac1d48	274	int fileno;
ff58fced	275
8812d7f2	276	ipo = calloc(1, sizeof(*ipo));
0d29de83	277	init_ipo(ipo);
26e616d0	278	fileno = trace_add_file(td, t->device, &bs);
ff58fced	279
24653680	280	ios[DDIR_TRIM]++;
26e616d0 JA	281	if (t->bytes > rw_bs[DDIR_TRIM])
26e616d0 JA	282	rw_bs[DDIR_TRIM] = t->bytes;
24653680	283
ff58fced JA	284	td->o.size += t->bytes;
ff58fced JA	285
ff58fced JA	286	INIT_FLIST_HEAD(&ipo->list);
ff58fced JA	287
26e616d0	288	ipo->offset = t->sector * bs;
0c63576e JA	289	if (td->o.replay_scale)
0c63576e JA	290	ipo->offset = ipo->offset / td->o.replay_scale;
a79f17bf	291	ipo_bytes_align(td->o.replay_align, ipo);
ff58fced JA	292	ipo->len = t->bytes;
	293	ipo->delay = ttime / 1000;
	294	ipo->ddir = DDIR_TRIM;
89ac1d48	295	ipo->fileno = fileno;
ff58fced JA	296
	297	dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
	298	ipo->offset, ipo->len,
	299	ipo->delay);
	300	queue_io_piece(td, ipo);
	301	}
	302
691c8fb0 JA	303	static void handle_trace_fs(struct thread_data td, struct blk_io_trace t,
691c8fb0 JA	304	unsigned long long ttime, unsigned long *ios,
26e616d0	305	unsigned int *rw_bs)
691c8fb0	306	{
26e616d0	307	unsigned int bs;
691c8fb0	308	int rw;
89ac1d48	309	int fileno;
5b3023b8	310
26e616d0	311	fileno = trace_add_file(td, t->device, &bs);
5b3023b8 JA	312
	313	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
	314
897c606a JA	315	/*
	316	* Need to figure out why 0 byte writes end up here sometimes, for
	317	* now just ignore them.
	318	*/
	319	if (!t->bytes)
	320	return;
	321	else if (t->bytes > rw_bs[rw])
26e616d0	322	rw_bs[rw] = t->bytes;
5b3023b8 JA	323
	324	ios[rw]++;
	325	td->o.size += t->bytes;
26e616d0	326	store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
cd991b9e JA	327	}
cd991b9e JA	328
65c42cc8 JA	329	static void handle_trace_flush(struct thread_data td, struct blk_io_trace t,
	330	unsigned long long ttime)
	331	{
	332	struct io_piece *ipo;
	333	unsigned int bs;
	334	int fileno;
	335
	336	ipo = calloc(1, sizeof(*ipo));
	337	init_ipo(ipo);
	338	fileno = trace_add_file(td, t->device, &bs);
	339
	340	ipo->delay = ttime / 1000;
	341	ipo->ddir = DDIR_SYNC;
	342	ipo->fileno = fileno;
	343
	344	dprint(FD_BLKTRACE, "store flush delay=%lu\n", ipo->delay);
	345	queue_io_piece(td, ipo);
	346	}
	347
691c8fb0 JA	348	/*
	349	* We only care for queue traces, most of the others are side effects
	350	* due to internal workings of the block layer.
	351	*/
	352	static void handle_trace(struct thread_data td, struct blk_io_trace t,
24653680	353	unsigned long ios, unsigned int bs)
691c8fb0	354	{
24653680	355	static unsigned long long last_ttime;
0c63576e	356	unsigned long long delay = 0;
24653680	357
691c8fb0 JA	358	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
691c8fb0 JA	359	return;
24653680 JA	360
24653680 JA	361	if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
6dd7fa77	362	if (!last_ttime \|\| td->o.no_stall)
24653680	363	delay = 0;
6dd7fa77	364	else if (td->o.replay_time_scale == 100)
24653680	365	delay = t->time - last_ttime;
6dd7fa77 JA	366	else {
	367	double tmp = t->time - last_ttime;
	368	double scale;
	369
	370	scale = (double) 100.0 / (double) td->o.replay_time_scale;
	371	tmp *= scale;
	372	delay = tmp;
24653680	373	}
6dd7fa77	374	last_ttime = t->time;
24653680	375	}
691c8fb0	376
0c63576e JA	377	t_bytes_align(&td->o, t);
0c63576e JA	378
691c8fb0	379	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
aec2de20	380	handle_trace_notify(t);
ff58fced	381	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
24653680	382	handle_trace_discard(td, t, delay, ios, bs);
65c42cc8 JA	383	else if (t->action & BLK_TC_ACT(BLK_TC_FLUSH))
65c42cc8 JA	384	handle_trace_flush(td, t, delay);
691c8fb0	385	else
24653680	386	handle_trace_fs(td, t, delay, ios, bs);
691c8fb0 JA	387	}
691c8fb0 JA	388
d95b34a6 JA	389	static void byteswap_trace(struct blk_io_trace *t)
	390	{
	391	t->magic = fio_swap32(t->magic);
	392	t->sequence = fio_swap32(t->sequence);
	393	t->time = fio_swap64(t->time);
	394	t->sector = fio_swap64(t->sector);
	395	t->bytes = fio_swap32(t->bytes);
	396	t->action = fio_swap32(t->action);
	397	t->pid = fio_swap32(t->pid);
	398	t->device = fio_swap32(t->device);
	399	t->cpu = fio_swap32(t->cpu);
	400	t->error = fio_swap16(t->error);
	401	t->pdu_len = fio_swap16(t->pdu_len);
	402	}
	403
24653680 JA	404	static int t_is_write(struct blk_io_trace *t)
	405	{
	406	return (t->action & BLK_TC_ACT(BLK_TC_WRITE \| BLK_TC_DISCARD)) != 0;
	407	}
	408
a6eaf6c9 JA	409	static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
	410	{
	411	if (t->action & BLK_TC_ACT(BLK_TC_READ))
	412	return DDIR_READ;
	413	else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
	414	return DDIR_WRITE;
	415	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
	416	return DDIR_TRIM;
	417
	418	return DDIR_INVAL;
	419	}
	420
	421	static void depth_inc(struct blk_io_trace t, int depth)
	422	{
	423	enum fio_ddir ddir;
	424
	425	ddir = t_get_ddir(t);
	426	if (ddir != DDIR_INVAL)
	427	depth[ddir]++;
	428	}
	429
	430	static void depth_dec(struct blk_io_trace t, int depth)
	431	{
	432	enum fio_ddir ddir;
	433
	434	ddir = t_get_ddir(t);
	435	if (ddir != DDIR_INVAL)
	436	depth[ddir]--;
	437	}
	438
	439	static void depth_end(struct blk_io_trace t, int this_depth, int *depth)
	440	{
	441	enum fio_ddir ddir = DDIR_INVAL;
	442
	443	ddir = t_get_ddir(t);
	444	if (ddir != DDIR_INVAL) {
	445	depth[ddir] = max(depth[ddir], this_depth[ddir]);
	446	this_depth[ddir] = 0;
	447	}
	448	}
	449
8c1fdf04 JA	450	/*
	451	* Load a blktrace file by reading all the blk_io_trace entries, and storing
	452	* them as io_pieces like the fio text version would do.
	453	*/
d95b34a6	454	int load_blktrace(struct thread_data td, const char filename, int need_swap)
fb7b71a3 JA	455	{
fb7b71a3 JA	456	struct blk_io_trace t;
24653680 JA	457	unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
24653680 JA	458	unsigned int rw_bs[DDIR_RWDIR_CNT];
e2887563	459	struct fifo *fifo;
5903e7b7	460	int fd, i, old_state;
89ac1d48	461	struct fio_file *f;
a6eaf6c9	462	int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
fb7b71a3 JA	463
	464	fd = open(filename, O_RDONLY);
	465	if (fd < 0) {
	466	td_verror(td, errno, "open blktrace file");
	467	return 1;
	468	}
	469
e2887563 JA	470	fifo = fifo_alloc(TRACE_FIFO_SIZE);
e2887563 JA	471
8edd973d	472	old_state = td_bump_runstate(td, TD_SETTING_UP);
5903e7b7	473
6df8adaa JA	474	td->o.size = 0;
6df8adaa JA	475
a6eaf6c9 JA	476	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
	477	ios[i] = 0;
	478	rw_bs[i] = 0;
	479	this_depth[i] = 0;
	480	depth[i] = 0;
	481	}
	482
4241ea8f	483	skipped_writes = 0;
fb7b71a3	484	do {
e2887563	485	int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
fb7b71a3	486
e2887563	487	if (ret < 0)
8c1fdf04	488	goto err;
e2887563 JA	489	else if (!ret)
	490	break;
	491	else if (ret < (int) sizeof(t)) {
	492	log_err("fio: short fifo get\n");
fb7b71a3	493	break;
fb7b71a3 JA	494	}
fb7b71a3 JA	495
d95b34a6 JA	496	if (need_swap)
	497	byteswap_trace(&t);
	498
fb7b71a3	499	if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
5ec10eaa JA	500	log_err("fio: bad magic in blktrace data: %x\n",
5ec10eaa JA	501	t.magic);
8c1fdf04	502	goto err;
fb7b71a3 JA	503	}
fb7b71a3 JA	504	if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
5ec10eaa JA	505	log_err("fio: bad blktrace version %d\n",
5ec10eaa JA	506	t.magic & 0xff);
8c1fdf04	507	goto err;
fb7b71a3	508	}
f12b323f JA	509	ret = discard_pdu(td, fifo, fd, &t);
f12b323f JA	510	if (ret < 0) {
fb7b71a3	511	td_verror(td, ret, "blktrace lseek");
8c1fdf04	512	goto err;
f12b323f JA	513	} else if (t.pdu_len != ret) {
	514	log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
	515	goto err;
fb7b71a3	516	}
691c8fb0	517	if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
eb5fdcf1	518	if ((t.action & 0xffff) == __BLK_TA_QUEUE)
a6eaf6c9 JA	519	depth_inc(&t, this_depth);
	520	else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) \|\|
	521	((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
	522	depth_dec(&t, this_depth);
	523	else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
	524	depth_end(&t, this_depth, depth);
691c8fb0	525
24653680	526	if (t_is_write(&t) && read_only) {
691c8fb0	527	skipped_writes++;
24653680	528	continue;
64bbb865	529	}
a6edd638	530	}
24653680 JA	531
24653680 JA	532	handle_trace(td, &t, ios, rw_bs);
fb7b71a3 JA	533	} while (1);
fb7b71a3 JA	534
3643a1f2	535	for_each_file(td, f, i)
89ac1d48	536	trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
89ac1d48	537
38470f85	538	fifo_free(fifo);
fb7b71a3	539	close(fd);
8c1fdf04	540
8edd973d	541	td_restore_runstate(td, old_state);
5903e7b7	542
f01b34ae JA	543	if (!td->files_index) {
	544	log_err("fio: did not find replay device(s)\n");
	545	return 1;
	546	}
	547
eb5fdcf1 JA	548	/*
	549	* For stacked devices, we don't always get a COMPLETE event so
	550	* the depth grows to insane values. Limit it to something sane(r).
	551	*/
a6eaf6c9 JA	552	max_depth = 0;
	553	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
	554	if (depth[i] > 1024)
	555	depth[i] = 1024;
	556	else if (!depth[i] && ios[i])
	557	depth[i] = 1;
	558	max_depth = max(depth[i], max_depth);
	559	}
eb5fdcf1	560
4241ea8f	561	if (skipped_writes)
5ec10eaa JA	562	log_err("fio: %s skips replay of %lu writes due to read-only\n",
5ec10eaa JA	563	td->o.name, skipped_writes);
4241ea8f	564
8c1fdf04 JA	565	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	566	log_err("fio: found no ios in blktrace data\n");
	567	return 1;
24653680	568	} else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
8c1fdf04	569	td->o.td_ddir = TD_DDIR_READ;
d84f8d49 JA	570	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
d84f8d49 JA	571	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
8c1fdf04	572	td->o.td_ddir = TD_DDIR_WRITE;
d84f8d49 JA	573	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
d84f8d49 JA	574	} else {
8c1fdf04	575	td->o.td_ddir = TD_DDIR_RW;
d84f8d49 JA	576	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
d84f8d49 JA	577	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
24653680	578	td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
d84f8d49	579	}
8c1fdf04 JA	580
	581	/*
	582	* We need to do direct/raw ios to the device, to avoid getting
76a00ec6 JA	583	* read-ahead in our way. But only do so if the minimum block size
76a00ec6 JA	584	* is a multiple of 4k, otherwise we don't know if it's safe to do so.
8c1fdf04	585	*/
76a00ec6	586	if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
a6eaf6c9	587	td->o.odirect = 1;
8c1fdf04	588
eb5fdcf1	589	/*
8a16f59b	590	* If depth wasn't manually set, use probed depth
eb5fdcf1	591	*/
8a16f59b	592	if (!fio_option_is_set(&td->o, iodepth))
a6eaf6c9	593	td->o.iodepth = td->o.iodepth_low = max_depth;
eb5fdcf1	594
fb7b71a3	595	return 0;
8c1fdf04 JA	596	err:
8c1fdf04 JA	597	close(fd);
38470f85	598	fifo_free(fifo);
8c1fdf04	599	return 1;
fb7b71a3	600	}