X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=c9a344f27aff5ebcb4a405a34e2da72d0b0b0d4a;hp=781599f48724c14530e81a0e6020c100fffd1a31;hb=1907dbc6c0c9e1b663156e64c64c172aaf04ae8a;hpb=755200a326a33e5e19b16dfd5e013dd98bcf1916

diff --git a/io_u.c b/io_u.c
index 781599f4..c9a344f2 100644
--- a/io_u.c
+++ b/io_u.c
@@ -3,13 +3,18 @@
 #include <string.h>
 #include <signal.h>
 #include <time.h>
+#include <assert.h>
 
 #include "fio.h"
 #include "os.h"
 
+/*
+ * Change this define to play with the timeout handling
+ */
+#undef FIO_USE_TIMEOUT
+
 struct io_completion_data {
 	int nr;				/* input */
-	endio_handler *handler;		/* input */
 
 	int error;			/* output */
 	unsigned long bytes_done[2];	/* output */
@@ -32,10 +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;
+	struct fio_file *f = io_u->file;
 	unsigned long long block;
 	unsigned int blocks;
 	unsigned int nr_blocks;
@@ -93,14 +98,14 @@ 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->sequential) {
+	if (td_random(td)) {
 		unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
 		int loops = 5;
 
@@ -129,9 +134,9 @@ static int get_next_offset(struct thread_data *td, struct fio_file *f,
 	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;
@@ -193,6 +198,9 @@ 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);
@@ -203,14 +211,16 @@ 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.
@@ -224,7 +234,6 @@ static int fill_io_u(struct thread_data *td, struct fio_file *f,
 	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;
 	}
 
@@ -234,18 +243,18 @@ 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->sequential && !td->norandommap)
-		mark_random_map(td, f, io_u);
+	if (!td->read_iolog && td_random(td) && !td->norandommap)
+		mark_random_map(td, io_u);
 
 	/*
 	 * If using a write iolog, store this entry.
@@ -253,7 +262,6 @@ static int fill_io_u(struct thread_data *td, struct fio_file *f,
 	if (td->write_iolog_file)
 		write_iolog_put(td, io_u);
 
-	io_u->file = f;
 	return 0;
 }
 
@@ -278,8 +286,8 @@ 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++;
 }
 
 static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
@@ -289,23 +297,23 @@ static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
 	switch (msec) {
 	default:
 		index++;
-	case 1024 ... 2047:
+	case 1000 ... 1999:
 		index++;
-	case 512 ... 1023:
+	case 750 ... 999:
 		index++;
-	case 256 ... 511:
+	case 500 ... 749:
 		index++;
-	case 128 ... 255:
+	case 250 ... 499:
 		index++;
-	case 64 ... 127:
+	case 100 ... 249:
 		index++;
-	case 32 ... 63:
+	case 50 ... 99:
 		index++;
-	case 16 ... 31:
+	case 20 ... 49:
 		index++;
-	case 8 ... 15:
+	case 10 ... 19:
 		index++;
-	case 4 ... 7:
+	case 4 ... 9:
 		index++;
 	case 2 ... 3:
 		index++;
@@ -313,7 +321,70 @@ 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)
+{
+	unsigned int fileno;
+	struct fio_file *f;
+
+	do {
+		long r = os_random_long(&td->next_file_state);
+
+		fileno = (unsigned int) ((double) (td->open_files * r) / (RAND_MAX + 1.0));
+		f = &td->files[fileno];
+		if (f->open)
+			return f;
+	} while (1);
+}
+
+/*
+ * 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)
+{
+	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->open_files)
+			td->next_file = 0;
+
+		if (f->open)
+			break;
+
+		f = NULL;
+	} while (td->next_file != old_next_file);
+
+	return f;
+}
+
+static struct fio_file *get_next_file(struct thread_data *td)
+{
+	struct fio_file *f;
+
+	if (!td->nr_open_files)
+		return NULL;
+
+	f = td->file_service_file;
+	if (f && f->open && td->file_service_left--)
+		return f;
+
+	if (td->file_service_type == FIO_FSERVICE_RR)
+		f = get_next_file_rr(td);
+	else
+		f = get_next_file_rand(td);
+
+	td->file_service_file = f;
+	td->file_service_left = td->file_service_nr - 1;
+	return f;
 }
 
 struct io_u *__get_io_u(struct thread_data *td)
@@ -328,9 +399,13 @@ 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) {
+		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);
@@ -345,8 +420,9 @@ 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);
@@ -357,18 +433,46 @@ struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
 	 * from a requeue, io_u already setup
 	 */
 	if (io_u->file)
-		return io_u;
+		goto out;
+
+	do {
+		f = get_next_file(td);
+		if (!f) {
+			put_io_u(td, io_u);
+			return NULL;
+		}
+
+		io_u->file = f;
+
+		if (!fill_io_u(td, io_u))
+			break;
+
+		/*
+		 * No more to do for this file, close it
+		 */
+		io_u->file = NULL;
+		td_io_close_file(td, f);
+
+		/*
+		 * probably not the right place to do this, but see
+		 * if we need to open a new file
+		 */
+		if (td->nr_open_files < td->nr_files &&
+		    td->open_files != td->nr_files) {
+			int ret = td_io_open_file(td, f);
+
+			if (ret) {
+				put_io_u(td, io_u);
+				return NULL;
+			}
+		}
+	} while (1);
 
 	if (td->zone_bytes >= td->zone_size) {
 		td->zone_bytes = 0;
 		f->last_pos += td->zone_skip;
 	}
 
-	if (fill_io_u(td, f, io_u)) {
-		put_io_u(td, io_u);
-		return NULL;
-	}
-
 	if (io_u->buflen + io_u->offset > f->real_file_size) {
 		if (td->io_ops->flags & FIO_RAWIO) {
 			put_io_u(td, io_u);
@@ -393,6 +497,7 @@ struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
 	/*
 	 * Set io data pointers.
 	 */
+out:
 	io_u->xfer_buf = io_u->buf;
 	io_u->xfer_buflen = io_u->buflen;
 
@@ -405,11 +510,31 @@ struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
 	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)
 {
 	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;
@@ -440,21 +565,21 @@ 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;
@@ -475,50 +600,142 @@ static void ios_completed(struct thread_data *td,
 	}
 }
 
-long io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
-			endio_handler *handler)
+/*
+ * Complete a single io_u for the sync engines.
+ */
+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;
 }
 
-long io_u_queued_complete(struct thread_data *td, int min_events,
-			  endio_handler *handler)
-
+/*
+ * Called to complete min_events number of io for the async engines.
+ */
+long io_u_queued_complete(struct thread_data *td, int min_events)
 {
-	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
-	struct timespec *tsp = NULL;
 	struct io_completion_data icd;
+	struct timespec *tvp = NULL;
 	int ret;
 
-	if (min_events > 0) {
-		tsp = &ts;
-		ret = td_io_commit(td);
-		if (ret < 0) {
-			td_verror(td, -ret);
-			return ret;
-		}
+	if (!min_events) {
+		struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
+
+		tvp = &ts;
 	}
 
-	ret = td_io_getevents(td, min_events, td->cur_depth, tsp);
+	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
 	if (ret < 0) {
-		td_verror(td, -ret);
+		td_verror(td, -ret, "td_io_getevents");
 		return ret;
 	} 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;
 }
+
+/*
+ * 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);
+}
+
+static void io_u_dump(struct io_u *io_u)
+{
+	unsigned long t_start = mtime_since_now(&io_u->start_time);
+	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
+
+	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
+	log_err("  buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, io_u->xfer_buf, io_u->buflen, io_u->xfer_buflen, io_u->offset);
+	log_err("  ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
+}
+#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();
+	struct list_head *entry;
+	struct io_u *io_u;
+	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);
+
+	list_for_each(entry, &td->io_u_busylist) {
+		io_u = list_entry(entry, struct io_u, list);
+
+		io_u_dump(io_u);
+	}
+
+	td_verror(td, ETIMEDOUT, "io_u timeout");
+	exit(1);
+}
+#endif
+
+void io_u_init_timeout(void)
+{
+#ifdef FIO_USE_TIMEOUT
+	signal(SIGALRM, io_u_timeout_handler);
+#endif
+}