X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=9cb68af4691025ff83f50c9d562292d94bfefccc;hp=48d407605738bc335ce7e9be2cf96f61a8c88979;hb=b370e46abdd2bfe764e28efb41c4caadc632b937;hpb=413dd459a7710ba421061e840dd9ac3161c70f20

diff --git a/io_u.c b/io_u.c
index 48d40760..9cb68af4 100644
--- a/io_u.c
+++ b/io_u.c
@@ -15,7 +15,6 @@
 
 struct io_completion_data {
 	int nr;				/* input */
-	endio_handler *handler;		/* input */
 
 	int error;			/* output */
 	unsigned long bytes_done[2];	/* output */
@@ -38,10 +37,10 @@ static int random_map_free(struct thread_data *td, struct fio_file *f,
 /*
  * Mark a given offset as used in the map.
  */
-static void mark_random_map(struct thread_data *td, struct fio_file *f,
-			    struct io_u *io_u)
+static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 {
-	unsigned int min_bs = td->rw_min_bs;
+	unsigned int min_bs = td->o.rw_min_bs;
+	struct fio_file *f = io_u->file;
 	unsigned long long block;
 	unsigned int blocks;
 	unsigned int nr_blocks;
@@ -80,7 +79,7 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
 
 	i = f->last_free_lookup;
 	*b = (i * BLOCKS_PER_MAP);
-	while ((*b) * td->rw_min_bs < f->real_file_size) {
+	while ((*b) * td->o.rw_min_bs < f->real_file_size) {
 		if (f->file_map[i] != -1UL) {
 			*b += ffz(f->file_map[i]);
 			f->last_free_lookup = i;
@@ -99,23 +98,26 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
  * until we find a free one. For sequential io, just return the end of
  * the last io issued.
  */
-static int get_next_offset(struct thread_data *td, struct fio_file *f,
-			   struct io_u *io_u)
+static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 {
+	struct fio_file *f = io_u->file;
 	const int ddir = io_u->ddir;
 	unsigned long long b, rb;
 	long r;
 
 	if (td_random(td)) {
-		unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
+		unsigned long long max_blocks = f->file_size / td->o.min_bs[ddir];
 		int loops = 5;
 
 		do {
 			r = os_random_long(&td->random_state);
-			b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
-			if (td->norandommap)
+			if (!max_blocks)
+				b = 0;
+			else
+				b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
+			if (td->o.norandommap)
 				break;
-			rb = b + (f->file_offset / td->min_bs[ddir]);
+			rb = b + (f->file_offset / td->o.min_bs[ddir]);
 			loops--;
 		} while (!random_map_free(td, f, rb) && loops);
 
@@ -126,41 +128,65 @@ static int get_next_offset(struct thread_data *td, struct fio_file *f,
 		if (!loops && get_next_free_block(td, f, &b))
 			return 1;
 	} else
-		b = f->last_pos / td->min_bs[ddir];
+		b = f->last_pos / td->o.min_bs[ddir];
 
-	io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
+	io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
 	if (io_u->offset >= f->real_file_size)
 		return 1;
 
 	return 0;
 }
 
-static unsigned int get_next_buflen(struct thread_data *td, struct fio_file *f,
-				    struct io_u *io_u)
+static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
 {
+	struct fio_file *f = io_u->file;
 	const int ddir = io_u->ddir;
 	unsigned int buflen;
 	long r;
 
-	if (td->min_bs[ddir] == td->max_bs[ddir])
-		buflen = td->min_bs[ddir];
+	if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
+		buflen = td->o.min_bs[ddir];
 	else {
 		r = os_random_long(&td->bsrange_state);
-		buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
-		if (!td->bs_unaligned)
-			buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
+		buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
+		if (!td->o.bs_unaligned)
+			buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
 	}
 
 	while (buflen + io_u->offset > f->real_file_size) {
-		if (buflen == td->min_bs[ddir])
+		if (buflen == td->o.min_bs[ddir]) {
+			if (!td->o.odirect) {
+				assert(io_u->offset <= f->real_file_size);
+				buflen = f->real_file_size - io_u->offset;
+				return buflen;
+			}
 			return 0;
+		}
 
-		buflen = td->min_bs[ddir];
+		buflen = td->o.min_bs[ddir];
 	}
 
 	return buflen;
 }
 
+static void set_rwmix_bytes(struct thread_data *td)
+{
+	unsigned long long rbytes;
+	unsigned int diff;
+
+	/*
+	 * we do time or byte based switch. this is needed because
+	 * buffered writes may issue a lot quicker than they complete,
+	 * whereas reads do not.
+	 */
+	rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
+	diff = td->o.rwmixread;
+	if (td->rwmix_ddir == DDIR_WRITE)
+		diff = 100 - diff;
+
+	td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * (100 - diff)) / diff;
+}
+
 /*
  * Return the data direction for the next io_u. If the job is a
  * mixed read/write workload, check the rwmix cycle and switch if
@@ -171,23 +197,34 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 	if (td_rw(td)) {
 		struct timeval now;
 		unsigned long elapsed;
+		unsigned int cycle;
 
 		fio_gettime(&now, NULL);
 	 	elapsed = mtime_since_now(&td->rwmix_switch);
 
+		cycle = td->o.rwmixcycle;
+		if (!td->rwmix_bytes)
+			cycle /= 10;
+
 		/*
 		 * Check if it's time to seed a new data direction.
 		 */
-		if (elapsed >= td->rwmixcycle) {
+		if (elapsed >= cycle &&
+		    td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
 			unsigned int v;
 			long r;
 
 			r = os_random_long(&td->rwmix_state);
 			v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
-			if (v < td->rwmixread)
+			if (v < td->o.rwmixread) {
+				if (td->rwmix_ddir != DDIR_READ)
+					set_rwmix_bytes(td);
 				td->rwmix_ddir = DDIR_READ;
-			else
+			} else {
+				if (td->rwmix_ddir != DDIR_WRITE)
+					set_rwmix_bytes(td);
 				td->rwmix_ddir = DDIR_WRITE;
+			}
 			memcpy(&td->rwmix_switch, &now, sizeof(now));
 		}
 		return td->rwmix_ddir;
@@ -212,28 +249,30 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 {
 	struct io_u *__io_u = *io_u;
 
+	__io_u->flags |= IO_U_F_FREE;
+	__io_u->flags &= ~IO_U_F_FLIGHT;
+
 	list_del(&__io_u->list);
 	list_add_tail(&__io_u->list, &td->io_u_requeues);
 	td->cur_depth--;
 	*io_u = NULL;
 }
 
-static int fill_io_u(struct thread_data *td, struct fio_file *f,
-		     struct io_u *io_u)
+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->read_iolog)
+	if (td->o.read_iolog)
 		return read_iolog_get(td, io_u);
 
 	/*
 	 * see if it's time to sync
 	 */
-	if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks)
-	    && td->io_issues[DDIR_WRITE] && should_fsync(td)) {
+	if (td->o.fsync_blocks &&
+	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
 		io_u->ddir = DDIR_SYNC;
-		io_u->file = f;
 		return 0;
 	}
 
@@ -243,33 +282,35 @@ static int fill_io_u(struct thread_data *td, struct fio_file *f,
 	 * No log, let the seq/rand engine retrieve the next buflen and
 	 * position.
 	 */
-	if (get_next_offset(td, f, io_u))
+	if (get_next_offset(td, io_u))
 		return 1;
 
-	io_u->buflen = get_next_buflen(td, f, io_u);
+	io_u->buflen = get_next_buflen(td, io_u);
 	if (!io_u->buflen)
 		return 1;
 
 	/*
 	 * mark entry before potentially trimming io_u
 	 */
-	if (!td->read_iolog && td_random(td) && !td->norandommap)
-		mark_random_map(td, f, io_u);
+	if (!td->o.read_iolog && td_random(td) && !td->o.norandommap)
+		mark_random_map(td, io_u);
 
 	/*
 	 * If using a write iolog, store this entry.
 	 */
-	if (td->write_iolog_file)
+	if (td->o.write_iolog_file)
 		write_iolog_put(td, io_u);
 
-	io_u->file = f;
 	return 0;
 }
 
-static void io_u_mark_depth(struct thread_data *td)
+void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
 {
 	int index = 0;
 
+	if (io_u->ddir == DDIR_SYNC)
+		return;
+
 	switch (td->cur_depth) {
 	default:
 		index++;
@@ -287,8 +328,8 @@ static void io_u_mark_depth(struct thread_data *td)
 		break;
 	}
 
-	td->io_u_map[index]++;
-	td->total_io_u++;
+	td->ts.io_u_map[index]++;
+	td->ts.total_io_u[io_u->ddir]++;
 }
 
 static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
@@ -322,22 +363,25 @@ static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
 		break;
 	}
 
-	td->io_u_lat[index]++;
+	td->ts.io_u_lat[index]++;
 }
 
 /*
  * Get next file to service by choosing one at random
  */
-static struct fio_file *get_next_file_rand(struct thread_data *td)
+static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
+					   int badf)
 {
-	long r = os_random_long(&td->next_file_state);
-	unsigned int fileno;
 	struct fio_file *f;
+	int fno;
 
 	do {
-		fileno = (unsigned int) ((double) (td->nr_files - 1) * r / (RAND_MAX + 1.0));
-		f = &td->files[fileno];
-		if (f->fd != -1)
+		long r = os_random_long(&td->next_file_state);
+
+		fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
+		f = &td->files[fno];
+
+		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
 			return f;
 	} while (1);
 }
@@ -345,7 +389,8 @@ static struct fio_file *get_next_file_rand(struct thread_data *td)
 /*
  * Get next file to service by doing round robin between all available ones
  */
-static struct fio_file *get_next_file_rr(struct thread_data *td)
+static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
+					 int badf)
 {
 	unsigned int old_next_file = td->next_file;
 	struct fio_file *f;
@@ -354,10 +399,10 @@ static struct fio_file *get_next_file_rr(struct thread_data *td)
 		f = &td->files[td->next_file];
 
 		td->next_file++;
-		if (td->next_file >= td->nr_files)
+		if (td->next_file >= td->o.nr_files)
 			td->next_file = 0;
 
-		if (f->fd != -1)
+		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
 			break;
 
 		f = NULL;
@@ -366,6 +411,41 @@ static struct fio_file *get_next_file_rr(struct thread_data *td)
 	return f;
 }
 
+static struct fio_file *get_next_file(struct thread_data *td)
+{
+	struct fio_file *f;
+
+	assert(td->o.nr_files <= td->files_index);
+
+	if (!td->nr_open_files)
+		return NULL;
+
+	f = td->file_service_file;
+	if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
+		return f;
+
+	if (td->o.file_service_type == FIO_FSERVICE_RR)
+		f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
+	else
+		f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
+
+	td->file_service_file = f;
+	td->file_service_left = td->file_service_nr - 1;
+	return f;
+}
+
+static struct fio_file *find_next_new_file(struct thread_data *td)
+{
+	struct fio_file *f;
+
+	if (td->o.file_service_type == FIO_FSERVICE_RR)
+		f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
+	else
+		f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
+
+	return f;
+}
+
 struct io_u *__get_io_u(struct thread_data *td)
 {
 	struct io_u *io_u = NULL;
@@ -378,6 +458,7 @@ struct io_u *__get_io_u(struct thread_data *td)
 		io_u->buflen = 0;
 		io_u->resid = 0;
 		io_u->file = NULL;
+		io_u->end_io = NULL;
 	}
 
 	if (io_u) {
@@ -388,7 +469,6 @@ struct io_u *__get_io_u(struct thread_data *td)
 		list_del(&io_u->list);
 		list_add(&io_u->list, &td->io_u_busylist);
 		td->cur_depth++;
-		io_u_mark_depth(td);
 	}
 
 	return io_u;
@@ -402,6 +482,7 @@ struct io_u *get_io_u(struct thread_data *td)
 {
 	struct fio_file *f;
 	struct io_u *io_u;
+	int ret;
 
 	io_u = __get_io_u(td);
 	if (!io_u)
@@ -413,26 +494,44 @@ struct io_u *get_io_u(struct thread_data *td)
 	if (io_u->file)
 		goto out;
 
-	if (td->file_service_type == FIO_FSERVICE_RR)
-		f = get_next_file_rr(td);
-	else
-		f = get_next_file_rand(td);
+	do {
+		f = get_next_file(td);
+		if (!f) {
+			put_io_u(td, io_u);
+			return NULL;
+		}
 
-	if (!f) {
-		put_io_u(td, io_u);
-		return NULL;
-	}
+set_file:
+		io_u->file = f;
 
-	io_u->file = f;
+		if (!fill_io_u(td, io_u))
+			break;
 
-	if (td->zone_bytes >= td->zone_size) {
-		td->zone_bytes = 0;
-		f->last_pos += td->zone_skip;
-	}
+		/*
+		 * No more to do for this file, close it
+		 */
+		io_u->file = NULL;
+		td_io_close_file(td, f);
 
-	if (fill_io_u(td, f, io_u)) {
-		put_io_u(td, io_u);
-		return NULL;
+		/*
+		 * probably not the right place to do this, but see
+		 * if we need to open a new file
+		 */
+		if (td->nr_open_files < td->o.open_files &&
+		    td->o.open_files != td->o.nr_files) {
+			f = find_next_new_file(td);
+
+			if (!f || (ret = td_io_open_file(td, f))) {
+				put_io_u(td, io_u);
+				return NULL;
+			}
+			goto set_file;
+		}
+	} while (1);
+
+	if (td->zone_bytes >= td->o.zone_size) {
+		td->zone_bytes = 0;
+		f->last_pos += td->o.zone_skip;
 	}
 
 	if (io_u->buflen + io_u->offset > f->real_file_size) {
@@ -452,7 +551,7 @@ struct io_u *get_io_u(struct thread_data *td)
 
 		f->last_pos = io_u->offset + io_u->buflen;
 
-		if (td->verify != VERIFY_NONE)
+		if (td->o.verify != VERIFY_NONE)
 			populate_verify_io_u(td, io_u);
 	}
 
@@ -472,6 +571,23 @@ out:
 	return io_u;
 }
 
+void io_u_log_error(struct thread_data *td, struct io_u *io_u)
+{
+	const char *msg[] = { "read", "write", "sync" };
+
+	log_err("fio: io_u error");
+
+	if (io_u->file)
+		log_err(" on file %s", io_u->file->file_name);
+
+	log_err(": %s\n", strerror(io_u->error));
+
+	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
+
+	if (!td->error)
+		td_verror(td, io_u->error, "io_u error");
+}
+
 static void io_completed(struct thread_data *td, struct io_u *io_u,
 			 struct io_completion_data *icd)
 {
@@ -480,6 +596,8 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 	assert(io_u->flags & IO_U_F_FLIGHT);
 	io_u->flags &= ~IO_U_F_FLIGHT;
 
+	put_file(td, io_u->file);
+
 	if (io_u->ddir == DDIR_SYNC) {
 		td->last_was_sync = 1;
 		return;
@@ -510,21 +628,21 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 
 		icd->bytes_done[idx] += bytes;
 
-		if (icd->handler) {
-			ret = icd->handler(io_u);
+		if (io_u->end_io) {
+			ret = io_u->end_io(io_u);
 			if (ret && !icd->error)
 				icd->error = ret;
 		}
-	} else
+	} else {
 		icd->error = io_u->error;
+		io_u_log_error(td, io_u);
+	}
 }
 
-static void init_icd(struct io_completion_data *icd, endio_handler *handler,
-		     int nr)
+static void init_icd(struct io_completion_data *icd, int nr)
 {
 	fio_gettime(&icd->time, NULL);
 
-	icd->handler = handler;
 	icd->nr = nr;
 
 	icd->error = 0;
@@ -548,39 +666,31 @@ static void ios_completed(struct thread_data *td,
 /*
  * Complete a single io_u for the sync engines.
  */
-long io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
-			endio_handler *handler)
+long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
 {
 	struct io_completion_data icd;
 
-	init_icd(&icd, handler, 1);
+	init_icd(&icd, 1);
 	io_completed(td, io_u, &icd);
 	put_io_u(td, io_u);
 
 	if (!icd.error)
 		return icd.bytes_done[0] + icd.bytes_done[1];
 
+	td_verror(td, icd.error, "io_u_sync_complete");
 	return -1;
 }
 
 /*
  * Called to complete min_events number of io for the async engines.
  */
-long io_u_queued_complete(struct thread_data *td, int min_events,
-			  endio_handler *handler)
-
+long io_u_queued_complete(struct thread_data *td, int min_events)
 {
 	struct io_completion_data icd;
 	struct timespec *tvp = NULL;
 	int ret;
 
-	if (min_events > 0) {
-		ret = td_io_commit(td);
-		if (ret < 0) {
-			td_verror(td, -ret, "td_io_commit");
-			return ret;
-		}
-	} else {
+	if (!min_events) {
 		struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
 		tvp = &ts;
@@ -593,11 +703,12 @@ long io_u_queued_complete(struct thread_data *td, int min_events,
 	} else if (!ret)
 		return ret;
 
-	init_icd(&icd, handler, ret);
+	init_icd(&icd, ret);
 	ios_completed(td, &icd);
 	if (!icd.error)
 		return icd.bytes_done[0] + icd.bytes_done[1];
 
+	td_verror(td, icd.error, "io_u_queued_complete");
 	return -1;
 }
 
@@ -672,7 +783,7 @@ static void io_u_timeout_handler(int fio_unused sig)
 		return;
 	}
 
-	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->name, td->pid);
+	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
 
 	list_for_each(entry, &td->io_u_busylist) {
 		io_u = list_entry(entry, struct io_u, list);