X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=fio.c;h=a394ff8c0d3045ad33e3fdc24f7e022a3e13bed0;hp=78deaf66a732b6e92434cf0ccef52b66b62d5a71;hb=50070b5a3b37f013a2b0b968e1b8b34bffff8589;hpb=97601024a53586d77a368763f08be9ee483fdc9c

diff --git a/fio.c b/fio.c
index 78deaf66..a394ff8c 100644
--- a/fio.c
+++ b/fio.c
@@ -136,27 +136,6 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 	return 0;
 }
 
-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;
-}
-
 /*
  * When job exits, we can cancel the in-flight IO if we are using async
  * io. Attempt to do so.
@@ -170,7 +149,9 @@ static void cleanup_pending_aio(struct thread_data *td)
 	/*
 	 * get immediately available events, if any
 	 */
-	io_u_queued_complete(td, 0, NULL);
+	r = io_u_queued_complete(td, 0, NULL);
+	if (r < 0)
+		return;
 
 	/*
 	 * now cancel remaining active events
@@ -179,14 +160,24 @@ static void cleanup_pending_aio(struct thread_data *td)
 		list_for_each_safe(entry, n, &td->io_u_busylist) {
 			io_u = list_entry(entry, struct io_u, list);
 
-			r = td->io_ops->cancel(td, io_u);
-			if (!r)
+			/*
+			 * if the io_u isn't in flight, then that generally
+			 * means someone leaked an io_u. complain but fix
+			 * it up, so we don't stall here.
+			 */
+			if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
+				log_err("fio: non-busy IO on busy list\n");
 				put_io_u(td, io_u);
+			} else {
+				r = td->io_ops->cancel(td, io_u);
+				if (!r)
+					put_io_u(td, io_u);
+			}
 		}
 	}
 
 	if (td->cur_depth)
-		io_u_queued_complete(td, td->cur_depth, NULL);
+		r = io_u_queued_complete(td, td->cur_depth, NULL);
 }
 
 /*
@@ -209,13 +200,14 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 		return 1;
 	}
 
+requeue:
 	ret = td_io_queue(td, io_u);
 	if (ret < 0) {
 		td_verror(td, io_u->error);
 		put_io_u(td, io_u);
 		return 1;
 	} else if (ret == FIO_Q_QUEUED) {
-		if (io_u_queued_complete(td, 1, NULL))
+		if (io_u_queued_complete(td, 1, NULL) < 0)
 			return 1;
 	} else if (ret == FIO_Q_COMPLETED) {
 		if (io_u->error) {
@@ -223,7 +215,12 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 			return 1;
 		}
 
-		io_u_sync_complete(td, io_u, NULL);
+		if (io_u_sync_complete(td, io_u, NULL) < 0)
+			return 1;
+	} else if (ret == FIO_Q_BUSY) {
+		if (td_io_commit(td))
+			return 1;
+		goto requeue;
 	}
 
 	return 0;
@@ -244,10 +241,15 @@ static void do_verify(struct thread_data *td)
 	 * read from disk.
 	 */
 	for_each_file(td, f, i) {
-		fio_io_sync(td, f);
-		file_invalidate_cache(td, f);
+		if (fio_io_sync(td, f))
+			break;
+		if (file_invalidate_cache(td, f))
+			break;
 	}
 
+	if (td->error)
+		return;
+
 	td_set_runstate(td, TD_VERIFYING);
 
 	io_u = NULL;
@@ -256,15 +258,20 @@ static void do_verify(struct thread_data *td)
 		if (!io_u)
 			break;
 
-		if (runtime_exceeded(td, &io_u->start_time))
+		if (runtime_exceeded(td, &io_u->start_time)) {
+			put_io_u(td, io_u);
 			break;
+		}
 
-		if (get_next_verify(td, io_u))
+		if (get_next_verify(td, io_u)) {
+			put_io_u(td, io_u);
 			break;
+		}
 
-		if (td_io_prep(td, io_u))
+		if (td_io_prep(td, io_u)) {
+			put_io_u(td, io_u);
 			break;
-
+		}
 requeue:
 		ret = td_io_queue(td, io_u);
 
@@ -279,18 +286,23 @@ requeue:
 				io_u->xfer_buf += bytes;
 				goto requeue;
 			}
-			if (io_u_sync_complete(td, io_u, verify_io_u))
+			ret = io_u_sync_complete(td, io_u, verify_io_u);
+			if (ret < 0)
 				break;
 			continue;
 		case FIO_Q_QUEUED:
 			break;
+		case FIO_Q_BUSY:
+			requeue_io_u(td, &io_u);
+			ret = td_io_commit(td);
+			break;
 		default:
 			assert(ret < 0);
 			td_verror(td, -ret);
 			break;
 		}
 
-		if (ret < 0)
+		if (ret < 0 || td->error)
 			break;
 
 		/*
@@ -298,20 +310,21 @@ requeue:
 		 * completed io_u's first.
 		 */
 		min_events = 0;
-		if (queue_full(td))
+		if (queue_full(td) || ret == FIO_Q_BUSY) {
 			min_events = 1;
 
+			if (td->cur_depth > td->iodepth_low)
+				min_events = td->cur_depth - td->iodepth_low;
+		}
+
 		/*
 		 * Reap required number of io units, if any, and do the
 		 * verification on them through the callback handler
 		 */
-		if (io_u_queued_complete(td, min_events, verify_io_u))
+		if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
 			break;
 	}
 
-	if (io_u)
-		put_io_u(td, io_u);
-
 	if (td->cur_depth)
 		cleanup_pending_aio(td);
 
@@ -351,7 +364,6 @@ static void do_io(struct thread_data *td)
 {
 	struct timeval s;
 	unsigned long usec;
-	struct fio_file *f;
 	int i, ret = 0;
 
 	td_set_runstate(td, TD_RUNNING);
@@ -365,11 +377,7 @@ static void do_io(struct thread_data *td)
 		if (td->terminate)
 			break;
 
-		f = get_next_file(td);
-		if (!f)
-			break;
-
-		io_u = get_io_u(td, f);
+		io_u = get_io_u(td);
 		if (!io_u)
 			break;
 
@@ -397,8 +405,21 @@ requeue:
 			}
 			fio_gettime(&comp_time, NULL);
 			bytes_done = io_u_sync_complete(td, io_u, NULL);
+			if (bytes_done < 0)
+				ret = bytes_done;
 			break;
 		case FIO_Q_QUEUED:
+			/*
+			 * if the engine doesn't have a commit hook,
+			 * the io_u is really queued. if it does have such
+			 * a hook, it has to call io_u_queued() itself.
+			 */
+			if (td->io_ops->commit == NULL)
+				io_u_queued(td, io_u);
+			break;
+		case FIO_Q_BUSY:
+			requeue_io_u(td, &io_u);
+			ret = td_io_commit(td);
 			break;
 		default:
 			assert(ret < 0);
@@ -406,19 +427,21 @@ requeue:
 			break;
 		}
 
-		if (ret < 0)
+		if (ret < 0 || td->error)
 			break;
 
-		add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
-
 		/*
 		 * See if we need to complete some commands
 		 */
-		if (ret == FIO_Q_QUEUED) {
+		if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 			min_evts = 0;
-			if (queue_full(td))
+			if (queue_full(td) || ret == FIO_Q_BUSY) {
 				min_evts = 1;
 
+				if (td->cur_depth > td->iodepth_low)
+					min_evts = td->cur_depth - td->iodepth_low;
+			}
+
 			fio_gettime(&comp_time, NULL);
 			bytes_done = io_u_queued_complete(td, min_evts, NULL);
 			if (bytes_done < 0)
@@ -462,6 +485,8 @@ requeue:
 	}
 
 	if (!td->error) {
+		struct fio_file *f;
+
 		if (td->cur_depth)
 			cleanup_pending_aio(td);
 
@@ -538,9 +563,12 @@ static int init_io_u(struct thread_data *td)
 			fill_rand_buf(io_u, max_bs);
 
 		io_u->index = i;
+		io_u->flags = IO_U_F_FREE;
 		list_add(&io_u->list, &td->io_u_freelist);
 	}
 
+	io_u_init_timeout();
+
 	return 0;
 }
 
@@ -600,11 +628,15 @@ static void clear_io_state(struct thread_data *td)
 	struct fio_file *f;
 	int i;
 
-	td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
+	td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 	td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 	td->zone_bytes = 0;
 
+	td->last_was_sync = 0;
+
 	for_each_file(td, f, i) {
+		f->last_completed_pos = 0;
+
 		f->last_pos = 0;
 		if (td->io_ops->flags & FIO_SYNCIO)
 			lseek(f->fd, SEEK_SET, 0);
@@ -630,6 +662,7 @@ static void *thread_main(void *data)
 
 	INIT_LIST_HEAD(&td->io_u_freelist);
 	INIT_LIST_HEAD(&td->io_u_busylist);
+	INIT_LIST_HEAD(&td->io_u_requeues);
 	INIT_LIST_HEAD(&td->io_hist_list);
 	INIT_LIST_HEAD(&td->io_log_list);
 
@@ -684,15 +717,16 @@ static void *thread_main(void *data)
 	}
 
 	fio_gettime(&td->epoch, NULL);
-	getrusage(RUSAGE_SELF, &td->ru_start);
+	memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
+	getrusage(RUSAGE_SELF, &td->ts.ru_start);
 
 	runtime[0] = runtime[1] = 0;
 	while (td->loops--) {
 		fio_gettime(&td->start, NULL);
-		memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
+		memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 
 		if (td->ratemin)
-			memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
+			memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 
 		clear_io_state(td);
 		prune_io_piece_log(td);
@@ -728,12 +762,12 @@ static void *thread_main(void *data)
 	td->runtime[0] = runtime[0] / 1000;
 	td->runtime[1] = runtime[1] / 1000;
 
-	if (td->bw_log)
-		finish_log(td, td->bw_log, "bw");
-	if (td->slat_log)
-		finish_log(td, td->slat_log, "slat");
-	if (td->clat_log)
-		finish_log(td, td->clat_log, "clat");
+	if (td->ts.bw_log)
+		finish_log(td, td->ts.bw_log, "bw");
+	if (td->ts.slat_log)
+		finish_log(td, td->ts.slat_log, "slat");
+	if (td->ts.clat_log)
+		finish_log(td, td->ts.clat_log, "clat");
 	if (td->write_iolog_file)
 		write_iolog_close(td);
 	if (td->exec_postrun) {
@@ -751,7 +785,7 @@ err:
 	close_ioengine(td);
 	cleanup_io_u(td);
 	td_set_runstate(td, TD_EXITED);
-	return (void *) td->error;
+	return (void *) (unsigned long) td->error;
 }
 
 /*
@@ -774,7 +808,7 @@ static int fork_main(int shmid, int offset)
 	td = data + offset * sizeof(struct thread_data);
 	ret = thread_main(td);
 	shmdt(data);
-	return (int) ret;
+	return (int) (unsigned long) ret;
 }
 
 /*
@@ -790,6 +824,8 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 	 */
 	pending = cputhreads = 0;
 	for_each_td(td, i) {
+		int flags;
+
 		/*
 		 * ->io_ops is NULL for a thread that has closed its
 		 * io engine
@@ -797,55 +833,57 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 		if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 			cputhreads++;
 
-		if (td->runstate < TD_EXITED) {
-			/*
-			 * check if someone quit or got killed in an unusual way
-			 */
-			ret = waitpid(td->pid, &status, WNOHANG);
-			if (ret < 0)
-				perror("waitpid");
-			else if ((ret == td->pid) && WIFSIGNALED(status)) {
+		if (!td->pid || td->runstate == TD_REAPED)
+			continue;
+
+		flags = WNOHANG;
+		if (td->runstate == TD_EXITED)
+			flags = 0;
+
+		/*
+		 * check if someone quit or got killed in an unusual way
+		 */
+		ret = waitpid(td->pid, &status, flags);
+		if (ret < 0 && !td->use_thread) {
+			if (errno == ECHILD) {
+				log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
+				td_set_runstate(td, TD_REAPED);
+				goto reaped;
+			}
+			perror("waitpid");
+		} else if (ret == td->pid) {
+			if (WIFSIGNALED(status)) {
 				int sig = WTERMSIG(status);
 
 				log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 				td_set_runstate(td, TD_REAPED);
 				goto reaped;
 			}
-		}
-
-		if (td->runstate != TD_EXITED) {
-			if (td->runstate < TD_RUNNING)
-				pending++;
-
-			continue;
-		}
-
-		if (td->error)
-			exit_value++;
-
-		td_set_runstate(td, TD_REAPED);
-
-		if (td->use_thread) {
-			long ret;
-
-			if (pthread_join(td->thread, (void *) &ret))
-				perror("thread_join");
-		} else {
-			int status;
-
-			ret = waitpid(td->pid, &status, 0);
-			if (ret < 0)
-				perror("waitpid");
-			else if (WIFEXITED(status) && WEXITSTATUS(status)) {
-				if (!exit_value)
-					exit_value++;
+			if (WIFEXITED(status)) {
+				if (WEXITSTATUS(status) && !td->error)
+					td->error = WEXITSTATUS(status);
+				if (td->use_thread) {
+					long ret;
+
+					if (pthread_join(td->thread, (void *) &ret))
+						perror("pthread_join");
+				}
+				td_set_runstate(td, TD_REAPED);
+				goto reaped;
 			}
 		}
 
+		/*
+		 * thread is not dead, continue
+		 */
+		continue;
 reaped:
 		(*nr_running)--;
 		(*m_rate) -= td->ratemin;
 		(*t_rate) -= td->rate;
+
+		if (td->error)
+			exit_value++;
 	}
 
 	if (*nr_running == cputhreads && !pending)
@@ -901,6 +939,8 @@ static void run_threads(void)
 		init_disk_util(td);
 	}
 
+	set_genesis_time();
+
 	while (todo) {
 		struct thread_data *map[MAX_JOBS];
 		struct timeval this_start;
@@ -1059,6 +1099,8 @@ int main(int argc, char *argv[])
 		setup_log(&agg_io_log[DDIR_WRITE]);
 	}
 
+	set_genesis_time();
+
 	disk_util_timer_arm();
 
 	run_threads();