X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=c93896001296419f1a348f2b5f9f46f660b0a366;hp=34c7e6f44bce50fed649380d391977de9752b32c;hb=55bc972818cdec46f47633da62758e4dcdb6f3f6;hpb=c685b5b2912bfbd3cdc8729d16a98438b1ff7593

diff --git a/io_u.c b/io_u.c
index 34c7e6f4..c9389600 100644
--- a/io_u.c
+++ b/io_u.c
@@ -8,6 +8,15 @@
 #include "fio.h"
 #include "os.h"
 
+struct io_completion_data {
+	int nr;				/* input */
+	endio_handler *handler;		/* input */
+
+	int error;			/* output */
+	unsigned long bytes_done[2];	/* output */
+	struct timeval time;		/* output */
+};
+
 /*
  * The ->file_map[] contains a map of blocks we have or have not done io
  * to yet. Used to make sure we cover the entire range in a fair fashion.
@@ -45,7 +54,7 @@ static void mark_random_map(struct thread_data *td, struct fio_file *f,
 		idx = RAND_MAP_IDX(td, f, block);
 		bit = RAND_MAP_BIT(td, f, block);
 
-		assert(idx < f->num_maps);
+		fio_assert(td, idx < f->num_maps);
 
 		f->file_map[idx] |= (1UL << bit);
 		block++;
@@ -105,24 +114,26 @@ static int get_next_offset(struct thread_data *td, struct fio_file *f,
 			loops--;
 		} while (!random_map_free(td, f, rb) && loops);
 
-		if (!loops) {
-			if (get_next_free_block(td, f, &b))
-				return 1;
-
-			b += f->file_offset / td->min_bs[ddir];
-		}
+		/*
+		 * if we failed to retrieve a truly random offset within
+		 * the loops assigned, see if there are free ones left at all
+		 */
+		if (!loops && get_next_free_block(td, f, &b))
+			return 1;
 	} else
 		b = f->last_pos / td->min_bs[ddir];
 
 	io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
-	if (io_u->offset > f->real_file_size)
+	if (io_u->offset >= f->real_file_size)
 		return 1;
 
 	return 0;
 }
 
-static unsigned int get_next_buflen(struct thread_data *td, int ddir)
+static unsigned int get_next_buflen(struct thread_data *td, struct fio_file *f,
+				    struct io_u *io_u)
 {
+	const int ddir = io_u->ddir;
 	unsigned int buflen;
 	long r;
 
@@ -135,15 +146,11 @@ static unsigned int get_next_buflen(struct thread_data *td, int ddir)
 			buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
 	}
 
-	if (buflen > td->io_size - td->this_io_bytes[ddir]) {
-		/*
-		 * if using direct/raw io, we may not be able to
-		 * shrink the size. so just fail it.
-		 */
-		if (td->io_ops->flags & FIO_RAWIO)
+	while (buflen + io_u->offset > f->real_file_size) {
+		if (buflen == td->min_bs[ddir])
 			return 0;
 
-		buflen = td->io_size - td->this_io_bytes[ddir];
+		buflen = td->min_bs[ddir];
 	}
 
 	return buflen;
@@ -187,12 +194,25 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 
 void put_io_u(struct thread_data *td, struct io_u *io_u)
 {
+	assert((io_u->flags & IO_U_F_FREE) == 0);
+	io_u->flags |= IO_U_F_FREE;
+
 	io_u->file = NULL;
 	list_del(&io_u->list);
 	list_add(&io_u->list, &td->io_u_freelist);
 	td->cur_depth--;
 }
 
+void requeue_io_u(struct thread_data *td, struct io_u **io_u)
+{
+	struct io_u *__io_u = *io_u;
+
+	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)
 {
@@ -205,8 +225,8 @@ static int fill_io_u(struct thread_data *td, struct fio_file *f,
 	/*
 	 * see if it's time to sync
 	 */
-	if (td->fsync_blocks && !(td->io_blocks[DDIR_WRITE] % td->fsync_blocks)
-	    && should_fsync(td)) {
+	if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks)
+	    && td->io_issues[DDIR_WRITE] && should_fsync(td)) {
 		io_u->ddir = DDIR_SYNC;
 		io_u->file = f;
 		return 0;
@@ -218,28 +238,27 @@ 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.
 	 */
-	io_u->buflen = get_next_buflen(td, io_u->ddir);
-	if (io_u->buflen) {
-		if (!get_next_offset(td, f, io_u)) {
-			/*
-			 * mark entry before potentially trimming io_u
-			 */
-			if (!td->read_iolog && !td->sequential &&
-			    !td->norandommap)
-				mark_random_map(td, f, io_u);
-
-			/*
-			 * If using a write iolog, store this entry.
-			 */
-			if (td->write_iolog_file)
-				write_iolog_put(td, io_u);
-
-			io_u->file = f;
-			return 0;
-		}
-	}
+	if (get_next_offset(td, f, io_u))
+		return 1;
 
-	return 1;
+	io_u->buflen = get_next_buflen(td, f, io_u);
+	if (!io_u->buflen)
+		return 1;
+
+	/*
+	 * mark entry before potentially trimming io_u
+	 */
+	if (!td->read_iolog && !td->sequential && !td->norandommap)
+		mark_random_map(td, f, io_u);
+
+	/*
+	 * If using a write iolog, store this entry.
+	 */
+	if (td->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)
@@ -267,16 +286,80 @@ static void io_u_mark_depth(struct thread_data *td)
 	td->total_io_u++;
 }
 
+static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
+{
+	int index = 0;
+
+	switch (msec) {
+	default:
+		index++;
+	case 1000 ... 1999:
+		index++;
+	case 750 ... 999:
+		index++;
+	case 500 ... 749:
+		index++;
+	case 250 ... 499:
+		index++;
+	case 100 ... 249:
+		index++;
+	case 50 ... 99:
+		index++;
+	case 20 ... 49:
+		index++;
+	case 10 ... 19:
+		index++;
+	case 4 ... 9:
+		index++;
+	case 2 ... 3:
+		index++;
+	case 0 ... 1:
+		break;
+	}
+
+	td->io_u_lat[index]++;
+}
+
+static struct fio_file *get_next_file(struct thread_data *td)
+{
+	unsigned int old_next_file = td->next_file;
+	struct fio_file *f;
+
+	do {
+		f = &td->files[td->next_file];
+
+		td->next_file++;
+		if (td->next_file >= td->nr_files)
+			td->next_file = 0;
+
+		if (f->fd != -1)
+			break;
+
+		f = NULL;
+	} while (td->next_file != old_next_file);
+
+	return f;
+}
+
 struct io_u *__get_io_u(struct thread_data *td)
 {
 	struct io_u *io_u = NULL;
 
-	if (!queue_full(td)) {
+	if (!list_empty(&td->io_u_requeues))
+		io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
+	else if (!queue_full(td)) {
 		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
 
 		io_u->buflen = 0;
-		io_u->error = 0;
 		io_u->resid = 0;
+		io_u->file = NULL;
+	}
+
+	if (io_u) {
+		assert(io_u->flags & IO_U_F_FREE);
+		io_u->flags &= ~IO_U_F_FREE;
+
+		io_u->error = 0;
 		list_del(&io_u->list);
 		list_add(&io_u->list, &td->io_u_busylist);
 		td->cur_depth++;
@@ -290,14 +373,29 @@ struct io_u *__get_io_u(struct thread_data *td)
  * Return an io_u to be processed. Gets a buflen and offset, sets direction,
  * etc. The returned io_u is fully ready to be prepped and submitted.
  */
-struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
+struct io_u *get_io_u(struct thread_data *td)
 {
+	struct fio_file *f;
 	struct io_u *io_u;
 
 	io_u = __get_io_u(td);
 	if (!io_u)
 		return NULL;
 
+	/*
+	 * from a requeue, io_u already setup
+	 */
+	if (io_u->file)
+		goto out;
+
+	f = get_next_file(td);
+	if (!f) {
+		put_io_u(td, io_u);
+		return NULL;
+	}
+
+	io_u->file = f;
+
 	if (td->zone_bytes >= td->zone_size) {
 		td->zone_bytes = 0;
 		f->last_pos += td->zone_skip;
@@ -323,32 +421,36 @@ struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
 			return NULL;
 		}
 
-		f->last_pos += io_u->buflen;
+		f->last_pos = io_u->offset + io_u->buflen;
 
 		if (td->verify != VERIFY_NONE)
 			populate_verify_io_u(td, io_u);
 	}
 
-	if (td_io_prep(td, io_u)) {
-		put_io_u(td, io_u);
-		return NULL;
-	}
-
 	/*
 	 * Set io data pointers.
 	 */
+out:
 	io_u->xfer_buf = io_u->buf;
 	io_u->xfer_buflen = io_u->buflen;
 
+	if (td_io_prep(td, io_u)) {
+		put_io_u(td, io_u);
+		return NULL;
+	}
+
 	fio_gettime(&io_u->start_time, NULL);
 	return io_u;
 }
 
-void io_completed(struct thread_data *td, struct io_u *io_u,
-		  struct io_completion_data *icd)
+static void io_completed(struct thread_data *td, struct io_u *io_u,
+			 struct io_completion_data *icd)
 {
 	unsigned long msec;
 
+	assert(io_u->flags & IO_U_F_FLIGHT);
+	io_u->flags &= ~IO_U_F_FLIGHT;
+
 	if (io_u->ddir == DDIR_SYNC) {
 		td->last_was_sync = 1;
 		return;
@@ -359,6 +461,7 @@ 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;
 		const enum fio_ddir idx = io_u->ddir;
+		int ret;
 
 		td->io_blocks[idx]++;
 		td->io_bytes[idx] += bytes;
@@ -371,24 +474,39 @@ void io_completed(struct thread_data *td, struct io_u *io_u,
 
 		add_clat_sample(td, idx, msec);
 		add_bw_sample(td, idx, &icd->time);
+		io_u_mark_latency(td, msec);
 
 		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
 			log_io_piece(td, io_u);
 
 		icd->bytes_done[idx] += bytes;
+
+		if (icd->handler) {
+			ret = icd->handler(io_u);
+			if (ret && !icd->error)
+				icd->error = ret;
+		}
 	} else
 		icd->error = io_u->error;
 }
 
-void ios_completed(struct thread_data *td, struct io_completion_data *icd)
+static void init_icd(struct io_completion_data *icd, endio_handler *handler,
+		     int nr)
 {
-	struct io_u *io_u;
-	int i;
-
 	fio_gettime(&icd->time, NULL);
 
+	icd->handler = handler;
+	icd->nr = nr;
+
 	icd->error = 0;
 	icd->bytes_done[0] = icd->bytes_done[1] = 0;
+}
+
+static void ios_completed(struct thread_data *td,
+			  struct io_completion_data *icd)
+{
+	struct io_u *io_u;
+	int i;
 
 	for (i = 0; i < icd->nr; i++) {
 		io_u = td->io_ops->event(td, i);
@@ -397,3 +515,131 @@ void ios_completed(struct thread_data *td, struct io_completion_data *icd)
 		put_io_u(td, io_u);
 	}
 }
+
+/*
+ * 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)
+{
+	struct io_completion_data icd;
+
+	init_icd(&icd, handler, 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];
+
+	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)
+
+{
+	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);
+			return ret;
+		}
+	} else {
+		struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
+
+		tvp = &ts;
+	}
+
+	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
+	if (ret < 0) {
+		td_verror(td, -ret);
+		return ret;
+	} else if (!ret)
+		return ret;
+
+	init_icd(&icd, handler, ret);
+	ios_completed(td, &icd);
+	if (!icd.error)
+		return icd.bytes_done[0] + icd.bytes_done[1];
+
+	return -1;
+}
+
+/*
+ * Call when io_u is really queued, to update the submission latency.
+ */
+void io_u_queued(struct thread_data *td, struct io_u *io_u)
+{
+	unsigned long slat_time;
+
+	slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
+	add_slat_sample(td, io_u->ddir, slat_time);
+}
+
+#ifdef FIO_USE_TIMEOUT
+void io_u_set_timeout(struct thread_data *td)
+{
+	assert(td->cur_depth);
+
+	td->timer.it_interval.tv_sec = 0;
+	td->timer.it_interval.tv_usec = 0;
+	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
+	td->timer.it_value.tv_usec = 0;
+	setitimer(ITIMER_REAL, &td->timer, NULL);
+	fio_gettime(&td->timeout_end, NULL);
+}
+#else
+void io_u_set_timeout(struct thread_data fio_unused *td)
+{
+}
+#endif
+
+#ifdef FIO_USE_TIMEOUT
+static void io_u_timeout_handler(int fio_unused sig)
+{
+	struct thread_data *td, *__td;
+	pid_t pid = getpid();
+	int i;
+
+	log_err("fio: io_u timeout\n");
+
+	/*
+	 * TLS would be nice...
+	 */
+	td = NULL;
+	for_each_td(__td, i) {
+		if (__td->pid == pid) {
+			td = __td;
+			break;
+		}
+	}
+
+	if (!td) {
+		log_err("fio: io_u timeout, can't find job\n");
+		exit(1);
+	}
+
+	if (!td->cur_depth) {
+		log_err("fio: timeout without pending work?\n");
+		return;
+	}
+
+	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->name, td->pid);
+	td->error = ETIMEDOUT;
+	exit(1);
+}
+#endif
+
+void io_u_init_timeout(void)
+{
+#ifdef FIO_USE_TIMEOUT
+	signal(SIGALRM, io_u_timeout_handler);
+#endif
+}