[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_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 (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
fb7b71a3 JA	1	/*
	2	* blktrace support code for fio
	3	*/
	4	#include <stdio.h>
	5	#include <stdlib.h>
5e6c2067	6	#include <sys/stat.h>
26e616d0 JA	7	#include <sys/ioctl.h>
26e616d0 JA	8	#include <linux/fs.h>
5e6c2067	9	#include <dirent.h>
8c1fdf04	10
01743ee1	11	#include "flist.h"
fb7b71a3 JA	12	#include "fio.h"
fb7b71a3 JA	13	#include "blktrace_api.h"
984f30c9	14	#include "oslib/linux-dev-lookup.h"
fb7b71a3	15
2da7df1d	16	#define TRACE_FIFO_SIZE 8192
e2887563 JA	17
	18	/*
	19	* fifo refill frontend, to avoid reading data in trace sized bites
	20	*/
	21	static int refill_fifo(struct thread_data td, struct fifo fifo, int fd)
	22	{
	23	char buf[TRACE_FIFO_SIZE];
f12b323f	24	unsigned int total;
e2887563 JA	25	int ret;
e2887563 JA	26
f12b323f JA	27	total = sizeof(buf);
	28	if (total > fifo_room(fifo))
	29	total = fifo_room(fifo);
e2887563	30
f12b323f JA	31	ret = read(fd, buf, total);
	32	if (ret < 0) {
	33	td_verror(td, errno, "read blktrace file");
	34	return -1;
e2887563 JA	35	}
e2887563 JA	36
f12b323f JA	37	if (ret > 0)
	38	ret = fifo_put(fifo, buf, ret);
	39
bd6f78b2	40	dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
f12b323f	41	return ret;
e2887563 JA	42	}
	43
	44	/*
	45	* Retrieve 'len' bytes from the fifo, refilling if necessary.
	46	*/
	47	static int trace_fifo_get(struct thread_data td, struct fifo fifo, int fd,
	48	void *buf, unsigned int len)
	49	{
f12b323f JA	50	if (fifo_len(fifo) < len) {
f12b323f JA	51	int ret = refill_fifo(td, fifo, fd);
e2887563	52
f12b323f JA	53	if (ret < 0)
	54	return ret;
	55	}
e2887563 JA	56
	57	return fifo_get(fifo, buf, len);
	58	}
	59
8c1fdf04 JA	60	/*
	61	* Just discard the pdu by seeking past it.
	62	*/
f12b323f JA	63	static int discard_pdu(struct thread_data td, struct fifo fifo, int fd,
f12b323f JA	64	struct blk_io_trace *t)
fb7b71a3 JA	65	{
	66	if (t->pdu_len == 0)
	67	return 0;
	68
bd6f78b2	69	dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
f12b323f	70	return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
fb7b71a3 JA	71	}
fb7b71a3 JA	72
8c1fdf04 JA	73	/*
	74	* Check if this is a blktrace binary data file. We read a single trace
	75	* into memory and check for the magic signature.
	76	*/
d95b34a6	77	int is_blktrace(const char filename, int need_swap)
fb7b71a3 JA	78	{
	79	struct blk_io_trace t;
	80	int fd, ret;
	81
	82	fd = open(filename, O_RDONLY);
4dced407	83	if (fd < 0)
fb7b71a3	84	return 0;
fb7b71a3 JA	85
	86	ret = read(fd, &t, sizeof(t));
	87	close(fd);
	88
	89	if (ret < 0) {
	90	perror("read blktrace");
	91	return 0;
	92	} else if (ret != sizeof(t)) {
	93	log_err("fio: short read on blktrace file\n");
	94	return 0;
	95	}
	96
d95b34a6 JA	97	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
d95b34a6 JA	98	*need_swap = 0;
fb7b71a3	99	return 1;
d95b34a6 JA	100	}
	101
	102	/*
	103	* Maybe it needs to be endian swapped...
	104	*/
	105	t.magic = fio_swap32(t.magic);
	106	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
	107	*need_swap = 1;
	108	return 1;
	109	}
fb7b71a3 JA	110
	111	return 0;
	112	}
	113
c69aa91f JA	114	#define FMINORBITS 20
	115	#define FMINORMASK ((1U << FMINORBITS) - 1)
	116	#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
	117	#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
eeb9c2aa	118
89ac1d48	119	static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
691c8fb0 JA	120	{
	121	struct io_piece *ipo;
	122
	123	ipo = calloc(1, sizeof(*ipo));
0d29de83	124	init_ipo(ipo);
691c8fb0 JA	125
	126	ipo->ddir = DDIR_INVAL;
	127	ipo->fileno = fileno;
89ac1d48	128	ipo->file_action = action;
01743ee1	129	flist_add_tail(&ipo->list, &td->io_log_list);
691c8fb0 JA	130	}
691c8fb0 JA	131
26e616d0	132	static int get_dev_blocksize(const char dev, unsigned int bs)
5e6c2067	133	{
26e616d0 JA	134	int fd;
	135
	136	fd = open(dev, O_RDONLY);
	137	if (fd < 0)
	138	return 1;
	139
	140	if (ioctl(fd, BLKSSZGET, bs) < 0) {
	141	close(fd);
	142	return 1;
	143	}
	144
	145	close(fd);
	146	return 0;
	147	}
	148
	149	static int trace_add_file(struct thread_data *td, __u32 device,
	150	unsigned int *bs)
	151	{
	152	static unsigned int last_maj, last_min, last_fileno, last_bs;
c69aa91f JA	153	unsigned int maj = FMAJOR(device);
c69aa91f JA	154	unsigned int min = FMINOR(device);
5e6c2067	155	struct fio_file *f;
5e6c2067	156	unsigned int i;
26e616d0	157	char dev[256];
5e6c2067	158
26e616d0 JA	159	if (last_maj == maj && last_min == min) {
26e616d0 JA	160	*bs = last_bs;
89ac1d48	161	return last_fileno;
26e616d0	162	}
5e6c2067 JA	163
	164	last_maj = maj;
	165	last_min = min;
	166
	167	/*
	168	* check for this file in our list
	169	*/
26e616d0	170	for_each_file(td, f, i) {
89ac1d48 SL	171	if (f->major == maj && f->minor == min) {
89ac1d48 SL	172	last_fileno = f->fileno;
26e616d0 JA	173	last_bs = f->bs;
26e616d0 JA	174	goto out;
89ac1d48	175	}
26e616d0	176	}
5e6c2067 JA	177
5e6c2067 JA	178	strcpy(dev, "/dev");
40bafa33	179	if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
26e616d0	180	unsigned int this_bs;
691c8fb0 JA	181	int fileno;
691c8fb0 JA	182
40bafa33 JA	183	if (td->o.replay_redirect)
	184	dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
	185	" with: %s\n", maj, min,
	186	td->o.replay_redirect);
	187	else
	188	dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
	189
bd6f78b2	190	dprint(FD_BLKTRACE, "add devices %s\n", dev);
49ffb4a2	191	fileno = add_file_exclusive(td, dev);
26e616d0 JA	192
	193	if (get_dev_blocksize(dev, &this_bs))
	194	this_bs = 512;
	195
b53f2c54	196	td->o.open_files++;
5903e7b7 JA	197	td->files[fileno]->major = maj;
5903e7b7 JA	198	td->files[fileno]->minor = min;
26e616d0	199	td->files[fileno]->bs = this_bs;
89ac1d48	200	trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
26e616d0	201
89ac1d48	202	last_fileno = fileno;
26e616d0	203	last_bs = this_bs;
bd6f78b2	204	}
f01b34ae	205
26e616d0 JA	206	out:
26e616d0 JA	207	*bs = last_bs;
89ac1d48	208	return last_fileno;
5e6c2067 JA	209	}
5e6c2067 JA	210
0c63576e JA	211	static void t_bytes_align(struct thread_options o, struct blk_io_trace t)
	212	{
	213	if (!o->replay_align)
	214	return;
	215
	216	t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
	217	}
	218
8c1fdf04 JA	219	/*
	220	* Store blk_io_trace data in an ipo for later retrieval.
	221	*/
fdefd987	222	static void store_ipo(struct thread_data *td, unsigned long long offset,
89ac1d48	223	unsigned int bytes, int rw, unsigned long long ttime,
26e616d0	224	int fileno, unsigned int bs)
fdefd987 JA	225	{
	226	struct io_piece ipo = malloc(sizeof(ipo));
	227
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 JA	271	{
ff58fced JA	272	struct io_piece ipo = malloc(sizeof(ipo));
26e616d0	273	unsigned int bs;
89ac1d48	274	int fileno;
ff58fced	275
0d29de83	276	init_ipo(ipo);
26e616d0	277	fileno = trace_add_file(td, t->device, &bs);
ff58fced	278
24653680	279	ios[DDIR_TRIM]++;
26e616d0 JA	280	if (t->bytes > rw_bs[DDIR_TRIM])
26e616d0 JA	281	rw_bs[DDIR_TRIM] = t->bytes;
24653680	282
ff58fced JA	283	td->o.size += t->bytes;
	284
	285	memset(ipo, 0, sizeof(*ipo));
	286	INIT_FLIST_HEAD(&ipo->list);
	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
26e616d0 JA	315	if (t->bytes > rw_bs[rw])
26e616d0 JA	316	rw_bs[rw] = t->bytes;
5b3023b8 JA	317
	318	ios[rw]++;
	319	td->o.size += t->bytes;
26e616d0	320	store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
cd991b9e JA	321	}
cd991b9e JA	322
691c8fb0 JA	323	/*
	324	* We only care for queue traces, most of the others are side effects
	325	* due to internal workings of the block layer.
	326	*/
	327	static void handle_trace(struct thread_data td, struct blk_io_trace t,
24653680	328	unsigned long ios, unsigned int bs)
691c8fb0	329	{
24653680	330	static unsigned long long last_ttime;
0c63576e	331	unsigned long long delay = 0;
24653680	332
691c8fb0 JA	333	if ((t->action & 0xffff) != __BLK_TA_QUEUE)
691c8fb0 JA	334	return;
24653680 JA	335
	336	if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
	337	if (!last_ttime \|\| td->o.no_stall) {
	338	last_ttime = t->time;
	339	delay = 0;
	340	} else {
	341	delay = t->time - last_ttime;
	342	last_ttime = t->time;
	343	}
	344	}
691c8fb0	345
0c63576e JA	346	t_bytes_align(&td->o, t);
0c63576e JA	347
691c8fb0	348	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
aec2de20	349	handle_trace_notify(t);
ff58fced	350	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
24653680	351	handle_trace_discard(td, t, delay, ios, bs);
691c8fb0	352	else
24653680	353	handle_trace_fs(td, t, delay, ios, bs);
691c8fb0 JA	354	}
691c8fb0 JA	355
d95b34a6 JA	356	static void byteswap_trace(struct blk_io_trace *t)
	357	{
	358	t->magic = fio_swap32(t->magic);
	359	t->sequence = fio_swap32(t->sequence);
	360	t->time = fio_swap64(t->time);
	361	t->sector = fio_swap64(t->sector);
	362	t->bytes = fio_swap32(t->bytes);
	363	t->action = fio_swap32(t->action);
	364	t->pid = fio_swap32(t->pid);
	365	t->device = fio_swap32(t->device);
	366	t->cpu = fio_swap32(t->cpu);
	367	t->error = fio_swap16(t->error);
	368	t->pdu_len = fio_swap16(t->pdu_len);
	369	}
	370
24653680 JA	371	static int t_is_write(struct blk_io_trace *t)
	372	{
	373	return (t->action & BLK_TC_ACT(BLK_TC_WRITE \| BLK_TC_DISCARD)) != 0;
	374	}
	375
a6eaf6c9 JA	376	static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
	377	{
	378	if (t->action & BLK_TC_ACT(BLK_TC_READ))
	379	return DDIR_READ;
	380	else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
	381	return DDIR_WRITE;
	382	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
	383	return DDIR_TRIM;
	384
	385	return DDIR_INVAL;
	386	}
	387
	388	static void depth_inc(struct blk_io_trace t, int depth)
	389	{
	390	enum fio_ddir ddir;
	391
	392	ddir = t_get_ddir(t);
	393	if (ddir != DDIR_INVAL)
	394	depth[ddir]++;
	395	}
	396
	397	static void depth_dec(struct blk_io_trace t, int depth)
	398	{
	399	enum fio_ddir ddir;
	400
	401	ddir = t_get_ddir(t);
	402	if (ddir != DDIR_INVAL)
	403	depth[ddir]--;
	404	}
	405
	406	static void depth_end(struct blk_io_trace t, int this_depth, int *depth)
	407	{
	408	enum fio_ddir ddir = DDIR_INVAL;
	409
	410	ddir = t_get_ddir(t);
	411	if (ddir != DDIR_INVAL) {
	412	depth[ddir] = max(depth[ddir], this_depth[ddir]);
	413	this_depth[ddir] = 0;
	414	}
	415	}
	416
8c1fdf04 JA	417	/*
	418	* Load a blktrace file by reading all the blk_io_trace entries, and storing
	419	* them as io_pieces like the fio text version would do.
	420	*/
d95b34a6	421	int load_blktrace(struct thread_data td, const char filename, int need_swap)
fb7b71a3 JA	422	{
fb7b71a3 JA	423	struct blk_io_trace t;
24653680 JA	424	unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
24653680 JA	425	unsigned int rw_bs[DDIR_RWDIR_CNT];
e2887563	426	struct fifo *fifo;
5903e7b7	427	int fd, i, old_state;
89ac1d48	428	struct fio_file *f;
a6eaf6c9	429	int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
fb7b71a3 JA	430
	431	fd = open(filename, O_RDONLY);
	432	if (fd < 0) {
	433	td_verror(td, errno, "open blktrace file");
	434	return 1;
	435	}
	436
e2887563 JA	437	fifo = fifo_alloc(TRACE_FIFO_SIZE);
e2887563 JA	438
8edd973d	439	old_state = td_bump_runstate(td, TD_SETTING_UP);
5903e7b7	440
6df8adaa JA	441	td->o.size = 0;
6df8adaa JA	442
a6eaf6c9 JA	443	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
	444	ios[i] = 0;
	445	rw_bs[i] = 0;
	446	this_depth[i] = 0;
	447	depth[i] = 0;
	448	}
	449
4241ea8f	450	skipped_writes = 0;
fb7b71a3	451	do {
e2887563	452	int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
fb7b71a3	453
e2887563	454	if (ret < 0)
8c1fdf04	455	goto err;
e2887563 JA	456	else if (!ret)
	457	break;
	458	else if (ret < (int) sizeof(t)) {
	459	log_err("fio: short fifo get\n");
fb7b71a3	460	break;
fb7b71a3 JA	461	}
fb7b71a3 JA	462
d95b34a6 JA	463	if (need_swap)
	464	byteswap_trace(&t);
	465
fb7b71a3	466	if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
5ec10eaa JA	467	log_err("fio: bad magic in blktrace data: %x\n",
5ec10eaa JA	468	t.magic);
8c1fdf04	469	goto err;
fb7b71a3 JA	470	}
fb7b71a3 JA	471	if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
5ec10eaa JA	472	log_err("fio: bad blktrace version %d\n",
5ec10eaa JA	473	t.magic & 0xff);
8c1fdf04	474	goto err;
fb7b71a3	475	}
f12b323f JA	476	ret = discard_pdu(td, fifo, fd, &t);
f12b323f JA	477	if (ret < 0) {
fb7b71a3	478	td_verror(td, ret, "blktrace lseek");
8c1fdf04	479	goto err;
f12b323f JA	480	} else if (t.pdu_len != ret) {
	481	log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
	482	goto err;
fb7b71a3	483	}
691c8fb0	484	if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
eb5fdcf1	485	if ((t.action & 0xffff) == __BLK_TA_QUEUE)
a6eaf6c9 JA	486	depth_inc(&t, this_depth);
	487	else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) \|\|
	488	((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
	489	depth_dec(&t, this_depth);
	490	else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
	491	depth_end(&t, this_depth, depth);
691c8fb0	492
24653680	493	if (t_is_write(&t) && read_only) {
691c8fb0	494	skipped_writes++;
24653680	495	continue;
64bbb865	496	}
a6edd638	497	}
24653680 JA	498
24653680 JA	499	handle_trace(td, &t, ios, rw_bs);
fb7b71a3 JA	500	} while (1);
fb7b71a3 JA	501
85a47ca2	502	for (i = 0; i < td->files_index; i++) {
f01b34ae	503	f = td->files[i];
89ac1d48	504	trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
85a47ca2	505	}
89ac1d48	506
38470f85	507	fifo_free(fifo);
fb7b71a3	508	close(fd);
8c1fdf04	509
8edd973d	510	td_restore_runstate(td, old_state);
5903e7b7	511
f01b34ae JA	512	if (!td->files_index) {
	513	log_err("fio: did not find replay device(s)\n");
	514	return 1;
	515	}
	516
eb5fdcf1 JA	517	/*
	518	* For stacked devices, we don't always get a COMPLETE event so
	519	* the depth grows to insane values. Limit it to something sane(r).
	520	*/
a6eaf6c9 JA	521	max_depth = 0;
	522	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
	523	if (depth[i] > 1024)
	524	depth[i] = 1024;
	525	else if (!depth[i] && ios[i])
	526	depth[i] = 1;
	527	max_depth = max(depth[i], max_depth);
	528	}
eb5fdcf1	529
4241ea8f	530	if (skipped_writes)
5ec10eaa JA	531	log_err("fio: %s skips replay of %lu writes due to read-only\n",
5ec10eaa JA	532	td->o.name, skipped_writes);
4241ea8f	533
8c1fdf04 JA	534	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
	535	log_err("fio: found no ios in blktrace data\n");
	536	return 1;
24653680	537	} else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
8c1fdf04	538	td->o.td_ddir = TD_DDIR_READ;
d84f8d49 JA	539	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
d84f8d49 JA	540	} else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
8c1fdf04	541	td->o.td_ddir = TD_DDIR_WRITE;
d84f8d49 JA	542	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
d84f8d49 JA	543	} else {
8c1fdf04	544	td->o.td_ddir = TD_DDIR_RW;
d84f8d49 JA	545	td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
d84f8d49 JA	546	td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
24653680	547	td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
d84f8d49	548	}
8c1fdf04 JA	549
	550	/*
	551	* We need to do direct/raw ios to the device, to avoid getting
76a00ec6 JA	552	* read-ahead in our way. But only do so if the minimum block size
76a00ec6 JA	553	* is a multiple of 4k, otherwise we don't know if it's safe to do so.
8c1fdf04	554	*/
76a00ec6	555	if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
a6eaf6c9	556	td->o.odirect = 1;
8c1fdf04	557
eb5fdcf1	558	/*
8a16f59b	559	* If depth wasn't manually set, use probed depth
eb5fdcf1	560	*/
8a16f59b	561	if (!fio_option_is_set(&td->o, iodepth))
a6eaf6c9	562	td->o.iodepth = td->o.iodepth_low = max_depth;
eb5fdcf1	563
fb7b71a3	564	return 0;
8c1fdf04 JA	565	err:
8c1fdf04 JA	566	close(fd);
38470f85	567	fifo_free(fifo);
8c1fdf04	568	return 1;
fb7b71a3	569	}