X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=fio.c;h=9510dee42971cf244a0cf1bdbec8548765c50d8c;hp=2fb7da9b76c22190e09e39efacd266572b29f63a;hb=8756e4d421722eaeb089067aeaaf317d05d53a57;hpb=838a3cd3b7587d3188dc0c75ec73ca068430f01b;ds=sidebyside

diff --git a/fio.c b/fio.c
index 2fb7da9b..9510dee4 100644
--- a/fio.c
+++ b/fio.c
@@ -69,7 +69,7 @@ enum {
 	TD_REAPED,
 };
 
-#define should_fsync(td)	(td_write(td) && (!(td)->odirect || (td)->override_sync))
+#define should_fsync(td)	((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
 
 static sem_t startup_sem;
 
@@ -541,6 +541,49 @@ static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
 	memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
 }
 
+unsigned int hweight32(unsigned int w)
+{
+	unsigned int res = w - ((w >> 1) & 0x55555555);
+
+	res = (res & 0x33333333) + ((res >> 2) & 0x33333333);
+	res = (res + (res >> 4)) & 0x0F0F0F0F;
+	res = res + (res >> 8);
+
+	return (res + (res >> 16)) & 0x000000FF;
+}
+
+unsigned long hweight64(unsigned long long w)
+{
+#if __WORDSIZE == 32
+        return hweight32((unsigned int)(w >> 32)) + hweight32((unsigned int)w);
+#elif __WORDSIZE == 64
+	unsigned long long v = w - ((w >> 1) & 0x5555555555555555ul);
+
+	v = (v & 0x3333333333333333ul) + ((v >> 2) & 0x3333333333333333ul);
+	v = (v + (v >> 4)) & 0x0F0F0F0F0F0F0F0Ful;
+	v = v + (v >> 8);
+	v = v + (v >> 16);
+
+	return (v + (v >> 32)) & 0x00000000000000FFul;
+#else
+#error __WORDSIZE not defined
+#endif
+}
+
+static int get_rw_ddir(struct thread_data *td)
+{
+	/*
+	 * perhaps cheasy, but use the hamming weight of the position
+	 * as a randomizer for data direction.
+	 */
+	if (td_rw(td))
+		return hweight64(td->last_pos) & 1;
+	else if (td_read(td))
+		return DDIR_READ;
+	else
+		return DDIR_WRITE;
+}
+
 /*
  * fill body of io_u->buf with random data and add a header with the
  * (eg) sha1sum of that data.
@@ -566,29 +609,59 @@ static void populate_io_u(struct thread_data *td, struct io_u *io_u)
 	memcpy(io_u->buf, &hdr, sizeof(hdr));
 }
 
-static int td_io_prep(struct thread_data *td, struct io_u *io_u, int read)
+static int td_io_prep(struct thread_data *td, struct io_u *io_u)
 {
-	if (read)
-		io_u->ddir = DDIR_READ;
-	else
-		io_u->ddir = DDIR_WRITE;
-
 	if (td->io_prep && td->io_prep(td, io_u))
 		return 1;
 
 	return 0;
 }
 
-static void put_io_u(struct thread_data *td, struct io_u *io_u)
+void put_io_u(struct thread_data *td, struct io_u *io_u)
 {
 	list_del(&io_u->list);
 	list_add(&io_u->list, &td->io_u_freelist);
 	td->cur_depth--;
 }
 
+static int fill_io_u(struct thread_data *td, struct io_u *io_u)
+{
+	/*
+	 * If using an iolog, grab next piece if any available.
+	 */
+	if (td->iolog) {
+		struct io_piece *ipo;
+
+		if (list_empty(&td->io_log_list))
+			return 1;
+
+		ipo = list_entry(td->io_log_list.next, struct io_piece, list);
+		list_del(&ipo->list);
+		io_u->offset = ipo->offset;
+		io_u->buflen = ipo->len;
+		io_u->ddir = ipo->ddir;
+		free(ipo);
+		return 0;
+	}
+
+	/*
+	 * No log, let the seq/rand engine retrieve the next position.
+	 */
+	if (!get_next_offset(td, &io_u->offset)) {
+		io_u->buflen = get_next_buflen(td);
+
+		if (io_u->buflen) {
+			io_u->ddir = get_rw_ddir(td);
+			return 0;
+		}
+	}
+
+	return 1;
+}
+
 #define queue_full(td)	(list_empty(&(td)->io_u_freelist))
 
-static struct io_u *__get_io_u(struct thread_data *td)
+struct io_u *__get_io_u(struct thread_data *td)
 {
 	struct io_u *io_u;
 
@@ -617,13 +690,7 @@ static struct io_u *get_io_u(struct thread_data *td)
 		td->last_pos += td->zone_skip;
 	}
 
-	if (get_next_offset(td, &io_u->offset)) {
-		put_io_u(td, io_u);
-		return NULL;
-	}
-
-	io_u->buflen = get_next_buflen(td);
-	if (!io_u->buflen) {
+	if (fill_io_u(td, io_u)) {
 		put_io_u(td, io_u);
 		return NULL;
 	}
@@ -636,7 +703,7 @@ static struct io_u *get_io_u(struct thread_data *td)
 		return NULL;
 	}
 
-	if (!td->sequential)
+	if (!td->iolog && !td->sequential)
 		mark_random_map(td, io_u);
 
 	td->last_pos += io_u->buflen;
@@ -644,7 +711,7 @@ static struct io_u *get_io_u(struct thread_data *td)
 	if (td->verify != VERIFY_NONE)
 		populate_io_u(td, io_u);
 
-	if (td_io_prep(td, io_u, td_read(td))) {
+	if (td_io_prep(td, io_u)) {
 		put_io_u(td, io_u);
 		return NULL;
 	}
@@ -659,8 +726,7 @@ static inline void td_set_runstate(struct thread_data *td, int runstate)
 	td->runstate = runstate;
 }
 
-static int get_next_verify(struct thread_data *td,
-			   unsigned long long *offset, unsigned int *len)
+static int get_next_verify(struct thread_data *td, struct io_u *io_u)
 {
 	struct io_piece *ipo;
 
@@ -670,8 +736,9 @@ static int get_next_verify(struct thread_data *td,
 	ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
 	list_del(&ipo->list);
 
-	*offset = ipo->offset;
-	*len = ipo->len;
+	io_u->offset = ipo->offset;
+	io_u->buflen = ipo->len;
+	io_u->ddir = DDIR_READ;
 	free(ipo);
 	return 0;
 }
@@ -724,6 +791,68 @@ static void log_io_piece(struct thread_data *td, struct io_u *io_u)
 	list_add(&ipo->list, entry);
 }
 
+static int init_iolog(struct thread_data *td)
+{
+	unsigned long long offset;
+	unsigned int bytes;
+	char *str, *p;
+	FILE *f;
+	int rw, i, reads, writes;
+
+	if (!td->iolog)
+		return 0;
+
+	f = fopen(td->iolog_file, "r");
+	if (!f) {
+		perror("fopen iolog");
+		return 1;
+	}
+
+	str = malloc(4096);
+	reads = writes = i = 0;
+	while ((p = fgets(str, 4096, f)) != NULL) {
+		struct io_piece *ipo;
+
+		if (sscanf(p, "%d,%llu,%u", &rw, &offset, &bytes) != 3) {
+			fprintf(stderr, "bad iolog: %s\n", p);
+			continue;
+		}
+		if (rw == DDIR_READ)
+			reads++;
+		else if (rw == DDIR_WRITE)
+			writes++;
+		else {
+			fprintf(stderr, "bad ddir: %d\n", rw);
+			continue;
+		}
+
+		ipo = malloc(sizeof(*ipo));
+		INIT_LIST_HEAD(&ipo->list);
+		ipo->offset = offset;
+		ipo->len = bytes;
+		if (bytes > td->max_bs)
+			td->max_bs = bytes;
+		ipo->ddir = rw;
+		list_add_tail(&ipo->list, &td->io_log_list);
+		i++;
+	}
+
+	free(str);
+	fclose(f);
+
+	if (!i)
+		return 1;
+
+	if (reads && !writes)
+		td->ddir = DDIR_READ;
+	else if (!reads && writes)
+		td->ddir = DDIR_READ;
+	else
+		td->iomix = 1;
+
+	return 0;
+}
+
 static int sync_td(struct thread_data *td)
 {
 	if (td->io_sync)
@@ -757,7 +886,7 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 
 	if (!io_u->error) {
 		unsigned int bytes = io_u->buflen - io_u->resid;
-		int idx = io_u->ddir;
+		const int idx = io_u->ddir;
 
 		td->io_blocks[idx]++;
 		td->io_bytes[idx] += bytes;
@@ -766,10 +895,10 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 
 		msec = mtime_since(&io_u->issue_time, &e);
 
-		add_clat_sample(td, io_u->ddir, msec);
-		add_bw_sample(td, io_u->ddir);
+		add_clat_sample(td, idx, msec);
+		add_bw_sample(td, idx);
 
-		if (td_write(td) && io_u->ddir == DDIR_WRITE)
+		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
 			log_io_piece(td, io_u);
 
 		icd->bytes_done[idx] += bytes;
@@ -867,12 +996,12 @@ static void do_verify(struct thread_data *td)
 		if (!io_u)
 			break;
 
-		if (get_next_verify(td, &io_u->offset, &io_u->buflen)) {
+		if (get_next_verify(td, io_u)) {
 			put_io_u(td, io_u);
 			break;
 		}
 
-		if (td_io_prep(td, io_u, 1)) {
+		if (td_io_prep(td, io_u)) {
 			put_io_u(td, io_u);
 			break;
 		}
@@ -1007,7 +1136,7 @@ static void do_io(struct thread_data *td)
 	if (td->cur_depth)
 		cleanup_pending_aio(td);
 
-	if (should_fsync(td))
+	if (should_fsync(td) && td->fsync_blocks)
 		sync_td(td);
 }
 
@@ -1029,6 +1158,8 @@ static int init_io(struct thread_data *td)
 		return fio_posixaio_init(td);
 	else if (td->io_engine == FIO_SGIO)
 		return fio_sgio_init(td);
+	else if (td->io_engine == FIO_SPLICEIO)
+		return fio_spliceio_init(td);
 	else {
 		fprintf(stderr, "bad io_engine %d\n", td->io_engine);
 		return 1;
@@ -1102,10 +1233,6 @@ static int init_io_u(struct thread_data *td)
 		}
 	}
 
-	INIT_LIST_HEAD(&td->io_u_freelist);
-	INIT_LIST_HEAD(&td->io_u_busylist);
-	INIT_LIST_HEAD(&td->io_hist_list);
-
 	p = ALIGN(td->orig_buffer);
 	for (i = 0; i < max_units; i++) {
 		io_u = malloc(sizeof(*io_u));
@@ -1113,6 +1240,7 @@ static int init_io_u(struct thread_data *td)
 		INIT_LIST_HEAD(&io_u->list);
 
 		io_u->buf = p + td->max_bs * i;
+		io_u->index = i;
 		list_add(&io_u->list, &td->io_u_freelist);
 	}
 
@@ -1208,6 +1336,7 @@ static int file_size(struct thread_data *td)
 	if (!td->file_size || td->file_size > td->real_file_size)
 		td->file_size = td->real_file_size;
 
+	td->file_size -= td->file_offset;
 	return 0;
 }
 
@@ -1230,23 +1359,26 @@ static int bdev_size(struct thread_data *td)
 	if (!td->file_size || td->file_size > td->real_file_size)
 		td->file_size = td->real_file_size;
 
+	td->file_size -= td->file_offset;
 	return 0;
 }
 
 static int get_file_size(struct thread_data *td)
 {
-	int ret;
+	int ret = 0;
 
 	if (td->filetype == FIO_TYPE_FILE)
 		ret = file_size(td);
-	else
+	else if (td->filetype == FIO_TYPE_BD)
 		ret = bdev_size(td);
+	else
+		td->real_file_size = -1;
 
 	if (ret)
 		return ret;
 
-	if (td->file_offset + td->file_size > td->real_file_size) {
-		fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset + td->file_size, td->real_file_size);
+	if (td->file_offset > td->real_file_size) {
+		fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->real_file_size);
 		return 1;
 	}
 
@@ -1268,14 +1400,15 @@ static int setup_file_mmap(struct thread_data *td)
 {
 	int flags;
 
-	if (td_read(td))
-		flags = PROT_READ;
-	else {
+	if (td_rw(td))
+		flags = PROT_READ | PROT_WRITE;
+	else if (td_write(td)) {
 		flags = PROT_WRITE;
 
 		if (td->verify != VERIFY_NONE)
 			flags |= PROT_READ;
-	}
+	} else
+		flags = PROT_READ;
 
 	td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
 	if (td->mmap == MAP_FAILED) {
@@ -1347,7 +1480,7 @@ static int setup_file(struct thread_data *td)
 		if (create_file(td, td->file_size, 0))
 			return 1;
 	} else if (td->filetype == FIO_TYPE_FILE) {
-		if (st.st_size < td->file_size) {
+		if (st.st_size < (off_t) td->file_size) {
 			if (create_file(td, td->file_size - st.st_size, 1))
 				return 1;
 		}
@@ -1356,9 +1489,7 @@ static int setup_file(struct thread_data *td)
 	if (td->odirect)
 		flags |= O_DIRECT;
 
-	if (td_read(td))
-		td->fd = open(td->file_name, flags | O_RDONLY);
-	else {
+	if (td_write(td) || td_rw(td)) {
 		if (td->filetype == FIO_TYPE_FILE) {
 			if (!td->overwrite)
 				flags |= O_TRUNC;
@@ -1371,6 +1502,13 @@ static int setup_file(struct thread_data *td)
 		flags |= O_RDWR;
 
 		td->fd = open(td->file_name, flags, 0600);
+	} else {
+		if (td->filetype == FIO_TYPE_CHAR)
+			flags |= O_RDWR;
+		else
+			flags |= O_RDONLY;
+
+		td->fd = open(td->file_name, flags);
 	}
 
 	if (td->fd == -1) {
@@ -1599,17 +1737,23 @@ static void init_disk_util(struct thread_data *td)
 		return;
 
 	/*
-	 * if this is inside a partition dir, jump back to parent
+	 * for md/dm, there's no queue dir. we already have the right place
 	 */
-	sprintf(tmp, "%s/queue", foo);
+	sprintf(tmp, "%s/stat", foo);
 	if (stat(tmp, &st)) {
-		p = dirname(foo);
-		sprintf(tmp, "%s/queue", p);
+		/*
+		 * if this is inside a partition dir, jump back to parent
+		 */
+		sprintf(tmp, "%s/queue", foo);
 		if (stat(tmp, &st)) {
-			fprintf(stderr, "unknown sysfs layout\n");
-			return;
+			p = dirname(foo);
+			sprintf(tmp, "%s/queue", p);
+			if (stat(tmp, &st)) {
+				fprintf(stderr, "unknown sysfs layout\n");
+				return;
+			}
+			sprintf(foo, "%s", p);
 		}
-		sprintf(foo, "%s", p);
 	}
 
 	disk_util_add(dev, foo);
@@ -1661,6 +1805,11 @@ static void *thread_main(void *data)
 
 	td->pid = getpid();
 
+	INIT_LIST_HEAD(&td->io_u_freelist);
+	INIT_LIST_HEAD(&td->io_u_busylist);
+	INIT_LIST_HEAD(&td->io_hist_list);
+	INIT_LIST_HEAD(&td->io_log_list);
+
 	if (init_io_u(td))
 		goto err;
 
@@ -1672,6 +1821,9 @@ static void *thread_main(void *data)
 	if (init_io(td))
 		goto err;
 
+	if (init_iolog(td))
+		goto err;
+
 	if (td->ioprio) {
 		if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 			td_verror(td, errno);
@@ -1704,6 +1856,9 @@ static void *thread_main(void *data)
 		do_io(td);
 
 		td->runtime[td->ddir] += mtime_since_now(&td->start);
+		if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
+			td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
+
 		update_rusage_stat(td);
 
 		if (td->error || td->terminate)
@@ -1828,7 +1983,8 @@ static void show_thread_status(struct thread_data *td,
 	printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
 
 	show_ddir_status(td, rs, td->ddir);
-	show_ddir_status(td, rs, td->ddir ^ 1);
+	if (td->io_bytes[td->ddir ^ 1])
+		show_ddir_status(td, rs, td->ddir ^ 1);
 
 	if (td->runtime[0] + td->runtime[1]) {
 		double runt = td->runtime[0] + td->runtime[1];
@@ -1858,7 +2014,12 @@ static void check_str_update(struct thread_data *td)
 			c = 'E';
 			break;
 		case TD_RUNNING:
-			if (td_read(td)) {
+			if (td_rw(td)) {
+				if (td->sequential)
+					c = 'M';
+				else
+					c = 'm';
+			} else if (td_read(td)) {
 				if (td->sequential)
 					c = 'R';
 				else
@@ -1915,14 +2076,80 @@ static void eta_to_str(char *str, int eta_sec)
 	str += sprintf(str, "%02ds", s);
 }
 
+static int thread_eta(struct thread_data *td, unsigned long elapsed)
+{
+	unsigned long long bytes_total, bytes_done;
+	unsigned int eta_sec = 0;
+
+	bytes_total = td->total_io_size;
+
+	/*
+	 * if writing, bytes_total will be twice the size. If mixing,
+	 * assume a 50/50 split and thus bytes_total will be 50% larger.
+	 */
+	if (td->verify) {
+		if (td_rw(td))
+			bytes_total = bytes_total * 3 / 2;
+		else
+			bytes_total <<= 1;
+	}
+	if (td->zone_size && td->zone_skip)
+		bytes_total /= (td->zone_skip / td->zone_size);
+
+	if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
+		double perc;
+
+		bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
+		perc = (double) bytes_done / (double) bytes_total;
+		if (perc > 1.0)
+			perc = 1.0;
+
+		eta_sec = (elapsed * (1.0 / perc)) - elapsed;
+
+		if (td->timeout && eta_sec > (td->timeout - elapsed))
+			eta_sec = td->timeout - elapsed;
+	} else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED) {
+		int t_eta = 0, r_eta = 0;
+
+		/*
+		 * We can only guess - assume it'll run the full timeout
+		 * if given, otherwise assume it'll run at the specified rate.
+		 */
+		if (td->timeout)
+			t_eta = td->timeout + td->start_delay - elapsed;
+		if (td->rate) {
+			r_eta = (bytes_total / 1024) / td->rate;
+			r_eta += td->start_delay - elapsed;
+		}
+
+		if (r_eta && t_eta)
+			eta_sec = min(r_eta, t_eta);
+		else if (r_eta)
+			eta_sec = r_eta;
+		else if (t_eta)
+			eta_sec = t_eta;
+		else
+			eta_sec = INT_MAX;
+	} else {
+		/*
+		 * thread is already done
+		 */
+		eta_sec = 0;
+	}
+
+	return eta_sec;
+}
+
 static void print_thread_status(void)
 {
-	unsigned long long bytes_done, bytes_total;
-	int i, nr_running, t_rate, m_rate, eta_sec;
+	unsigned long elapsed = time_since_now(&genesis);
+	int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
 	char eta_str[32];
-	double perc;
+	double perc = 0.0;
+
+	eta_secs = malloc(thread_number * sizeof(int));
+	memset(eta_secs, 0, thread_number * sizeof(int));
 
-	bytes_done = bytes_total = 0;
 	nr_running = t_rate = m_rate = 0;
 	for (i = 0; i < thread_number; i++) {
 		struct thread_data *td = &threads[i];
@@ -1933,45 +2160,44 @@ static void print_thread_status(void)
 			m_rate += td->ratemin;
 		}
 
-		bytes_total += td->total_io_size;
-		if (td->verify)
-			bytes_total += td->total_io_size;
-
-		if (td->zone_size && td->zone_skip)
-			bytes_total /= (td->zone_skip / td->zone_size);
-	
-		bytes_done += td->io_bytes[DDIR_READ] +td->io_bytes[DDIR_WRITE];
+		if (elapsed >= 3)
+			eta_secs[i] = thread_eta(td, elapsed);
+		else
+			eta_secs[i] = INT_MAX;
 
 		check_str_update(td);
 	}
 
-	perc = 0;
-	eta_sec = 0;
-	if (bytes_total && bytes_done) {
-		unsigned long runtime;
-
-		perc = (double) bytes_done / (double) bytes_total;
-		if (perc > 1.0)
-			perc = 1.0;
+	if (exitall_on_terminate)
+		eta_sec = INT_MAX;
+	else
+		eta_sec = 0;
 
-		runtime = time_since_now(&genesis);
-		if (runtime >= 5) {
-			memset(eta_str, 0, sizeof(eta_str));
-			eta_sec = (runtime * (1.0 / perc)) - runtime;
-			eta_to_str(eta_str, eta_sec);
+	for (i = 0; i < thread_number; i++) {
+		if (exitall_on_terminate) {
+			if (eta_secs[i] < eta_sec)
+				eta_sec = eta_secs[i];
+		} else {
+			if (eta_secs[i] > eta_sec)
+				eta_sec = eta_secs[i];
 		}
+	}
 
-		perc *= 100.0;
+	if (eta_sec != INT_MAX && elapsed) {
+		perc = (double) elapsed / (double) (elapsed + eta_sec);
+		eta_to_str(eta_str, eta_sec);
 	}
 
 	printf("Threads now running (%d)", nr_running);
 	if (m_rate || t_rate)
 		printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
-	printf(": [%s] [%3.2f%% done]", run_str, perc);
-	if (eta_sec)
-		printf(" [eta %s]", eta_str);
+	if (eta_sec != INT_MAX) {
+		perc *= 100.0;
+		printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
+	}
 	printf("\r");
 	fflush(stdout);
+	free(eta_secs);
 }
 
 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)