X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=blktrace.c;h=36a7180930d0f23bf036922820a7901f165a15c9;hp=8ba538ae8abc57dc1093df9273b2a894aae81cd8;hb=d22e2fc8834280e5ee789e2a1e2b23ebbbeabb7e;hpb=f12b323f5d5ca40cd966ebbcfbbfcdf0a1fc229e

diff --git a/blktrace.c b/blktrace.c
index 8ba538ae..36a71809 100644
--- a/blktrace.c
+++ b/blktrace.c
@@ -3,12 +3,16 @@
  */
 #include <stdio.h>
 #include <stdlib.h>
+#include <sys/ioctl.h>
+#include <linux/fs.h>
 
-#include "list.h"
+#include "flist.h"
 #include "fio.h"
+#include "blktrace.h"
 #include "blktrace_api.h"
+#include "oslib/linux-dev-lookup.h"
 
-#define TRACE_FIFO_SIZE	65536
+#define TRACE_FIFO_SIZE	8192
 
 /*
  * fifo refill frontend, to avoid reading data in trace sized bites
@@ -32,6 +36,7 @@ static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
 	if (ret > 0)
 		ret = fifo_put(fifo, buf, ret);
 
+	dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
 	return ret;
 }
 
@@ -60,6 +65,7 @@ static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
 	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);
 }
 
@@ -67,56 +73,296 @@ static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
  * 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)
+bool is_blktrace(const char *filename, int *need_swap)
 {
 	struct blk_io_trace t;
 	int fd, ret;
 
 	fd = open(filename, O_RDONLY);
-	if (fd < 0) {
-		perror("open blktrace");
-		return 0;
-	}
+	if (fd < 0)
+		return false;
 
 	ret = read(fd, &t, sizeof(t));
 	close(fd);
 
 	if (ret < 0) {
 		perror("read blktrace");
-		return 0;
+		return false;
 	} else if (ret != sizeof(t)) {
 		log_err("fio: short read on blktrace file\n");
-		return 0;
+		return false;
 	}
 
-	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
+	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
+		*need_swap = 0;
+		return true;
+	}
+
+	/*
+	 * 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 true;
+	}
+
+	return false;
+}
+
+#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)
+		      unsigned int bytes, int rw, unsigned long long ttime,
+		      int fileno, unsigned int bs)
 {
-	struct io_piece *ipo = malloc(sizeof(*ipo));
+	struct io_piece *ipo;
 
-	memset(ipo, 0, sizeof(*ipo));
-	INIT_LIST_HEAD(&ipo->list);
-	/*
-	 * the 512 is wrong here, it should be the hardware sector size...
-	 */
-	ipo->offset = offset * 512;
+	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;
+
+	if (td->o.replay_skip & (1u << DDIR_TRIM))
+		return;
+
+	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 dump_trace(struct blk_io_trace *t)
+{
+	log_err("blktrace: ignoring zero byte trace: action=%x\n", t->action);
+}
+
+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 (rw) {
+		if (td->o.replay_skip & (1u << DDIR_WRITE))
+			return;
+	} else {
+		if (td->o.replay_skip & (1u << DDIR_READ))
+			return;
+	}
+
+	if (!t->bytes) {
+		if (!fio_did_warn(FIO_WARN_BTRACE_ZERO))
+			dump_trace(t);
+		return;
+	}
+
+	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, unsigned long *ios)
+{
+	struct io_piece *ipo;
+	unsigned int bs;
+	int fileno;
+
+	if (td->o.replay_skip & (1u << DDIR_SYNC))
+		return;
+
+	ipo = calloc(1, sizeof(*ipo));
+	init_ipo(ipo);
+	fileno = trace_add_file(td, t->device, &bs);
 
-	list_add_tail(&ipo->list, &td->io_log_list);
+	ipo->delay = ttime / 1000;
+	ipo->ddir = DDIR_SYNC;
+	ipo->fileno = fileno;
+
+	ios[DDIR_SYNC]++;
+	dprint(FD_BLKTRACE, "store flush delay=%lu\n", ipo->delay);
+	queue_io_piece(td, ipo);
 }
 
 /*
@@ -124,60 +370,131 @@ static void store_ipo(struct thread_data *td, unsigned long long offset,
  * due to internal workings of the block layer.
  */
 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
-			 unsigned long long ttime, unsigned long *ios,
-			 unsigned int *bs)
+			 unsigned long *ios, unsigned int *bs)
 {
-	int rw;
+	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_PC))
-		return;
 
-	/*
-	 * should not happen, need to look into that...
-	 */
-	if (!t->bytes)
-		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;
+	}
 
-	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
+	t_bytes_align(&td->o, t);
 
-	if (t->bytes > bs[rw])
-		bs[rw] = t->bytes;
+	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, ios);
+	else
+		handle_trace_fs(td, t, delay, ios, bs);
+}
 
-	ios[rw]++;
-	td->o.size += t->bytes;
-	store_ipo(td, t->sector, t->bytes, rw, ttime);
+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 bool 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)
+bool load_blktrace(struct thread_data *td, const char *filename, int need_swap)
 {
-	unsigned long long ttime, delay;
 	struct blk_io_trace t;
-	unsigned long ios[2];
-	unsigned int cpu;
-	unsigned int rw_bs[2];
+	unsigned long ios[DDIR_RWDIR_SYNC_CNT] = { };
+	unsigned int rw_bs[DDIR_RWDIR_CNT] = { };
+	unsigned long skipped_writes;
 	struct fifo *fifo;
-	int fd;
+	int fd, i, old_state, max_depth;
+	struct fio_file *f;
+	int this_depth[DDIR_RWDIR_CNT] = { };
+	int depth[DDIR_RWDIR_CNT] = { };
 
 	fd = open(filename, O_RDONLY);
 	if (fd < 0) {
 		td_verror(td, errno, "open blktrace file");
-		return 1;
+		return false;
 	}
 
 	fifo = fifo_alloc(TRACE_FIFO_SIZE);
 
-	td->o.size = 0;
+	old_state = td_bump_runstate(td, TD_SETTING_UP);
 
-	cpu = 0;
-	ttime = 0;
-	ios[0] = ios[1] = 0;
-	rw_bs[0] = rw_bs[1] = 0;
+	td->o.size = 0;
+	skipped_writes = 0;
 	do {
 		int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
 
@@ -190,12 +507,17 @@ int load_blktrace(struct thread_data *td, const char *filename)
 			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);
+			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);
+			log_err("fio: bad blktrace version %d\n",
+								t.magic & 0xff);
 			goto err;
 		}
 		ret = discard_pdu(td, fifo, fd, &t);
@@ -206,25 +528,59 @@ int load_blktrace(struct thread_data *td, const char *filename)
 			log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
 			goto err;
 		}
-		if (!ttime) {
-			ttime = t.time;
-			cpu = t.cpu;
+		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;
+			}
 		}
-		delay = 0;
-		if (cpu == t.cpu)
-			delay = t.time - ttime;
-		handle_trace(td, &t, delay, ios, rw_bs);
-		ttime = t.time;
-		cpu = t.cpu;
+
+		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);
 
-	if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
+	td_restore_runstate(td, old_state);
+
+	if (!td->files_index) {
+		log_err("fio: did not find replay device(s)\n");
+		return false;
+	}
+
+	/*
+	 * 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] && !ios[DDIR_TRIM] &&
+	    !ios[DDIR_SYNC]) {
 		log_err("fio: found no ios in blktrace data\n");
-		return 1;
-	} else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
+		return false;
+	} 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]) {
@@ -234,17 +590,26 @@ int load_blktrace(struct thread_data *td, const char *filename)
 		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.
+	 * 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.
 	 */
-	td->o.odirect = 1;
+	if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
+		td->o.odirect = 1;
 
-	return 0;
+	/*
+	 * 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 true;
 err:
 	close(fd);
 	fifo_free(fifo);
-	return 1;
+	return false;
 }