X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=fio.c;h=504b78db0b0cd6dc1390435f56fd89d6bf19345c;hp=28f6376dfeac917305280324aafa822ff9e9e3f8;hb=4b87898e8d76aaf05baec83077a11311c1447397;hpb=8d7ad8d9d737982f7967f91f00ba9a56402fc60b

diff --git a/fio.c b/fio.c
index 28f6376d..504b78db 100644
--- a/fio.c
+++ b/fio.c
@@ -26,31 +26,43 @@
 #include <string.h>
 #include <signal.h>
 #include <time.h>
+#include <locale.h>
 #include <assert.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <sys/ipc.h>
 #include <sys/shm.h>
-#include <sys/ioctl.h>
 #include <sys/mman.h>
 
 #include "fio.h"
 #include "os.h"
 
-#define MASK	(4095)
-
-#define ALIGN(buf)	(char *) (((unsigned long) (buf) + MASK) & ~(MASK))
+unsigned long page_mask;
+unsigned long page_size;
+#define ALIGN(buf)	\
+	(char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
 
 int groupid = 0;
 int thread_number = 0;
+int nr_process = 0;
+int nr_thread = 0;
 int shm_id = 0;
 int temp_stall_ts;
 
-static volatile int startup_sem;
+static struct fio_sem *startup_sem;
+static volatile int fio_abort;
+static int exit_value;
+
+struct io_log *agg_io_log[2];
 
 #define TERMINATE_ALL		(-1)
 #define JOB_START_TIMEOUT	(5 * 1000)
 
+static inline void td_set_runstate(struct thread_data *td, int runstate)
+{
+	td->runstate = runstate;
+}
+
 static void terminate_threads(int group_id)
 {
 	struct thread_data *td;
@@ -58,8 +70,13 @@ static void terminate_threads(int group_id)
 
 	for_each_td(td, i) {
 		if (group_id == TERMINATE_ALL || groupid == td->groupid) {
+			/*
+			 * if the thread is running, just let it exit
+			 */
+			if (td->runstate < TD_RUNNING)
+				kill(td->pid, SIGQUIT);
 			td->terminate = 1;
-			td->start_delay = 0;
+			td->o.start_delay = 0;
 		}
 	}
 }
@@ -73,7 +90,7 @@ static void sig_handler(int sig)
 			print_thread_status();
 			break;
 		default:
-			printf("\nfio: terminating on signal\n");
+			printf("\nfio: terminating on signal %d\n", sig);
 			fflush(stdout);
 			terminate_threads(TERMINATE_ALL);
 			break;
@@ -85,9 +102,16 @@ static void sig_handler(int sig)
  */
 static int check_min_rate(struct thread_data *td, struct timeval *now)
 {
+	unsigned long long bytes = 0;
+	unsigned long iops = 0;
 	unsigned long spent;
 	unsigned long rate;
-	int ddir = td->ddir;
+
+	/*
+	 * No minimum rate set, always ok
+	 */
+	if (!td->o.ratemin && !td->o.rate_iops_min)
+		return 0;
 
 	/*
 	 * allow a 2 second settle period in the beginning
@@ -95,82 +119,85 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
 	if (mtime_since(&td->start, now) < 2000)
 		return 0;
 
+	if (td_read(td)) {
+		iops += td->io_blocks[DDIR_READ];
+		bytes += td->this_io_bytes[DDIR_READ];
+	}
+	if (td_write(td)) {
+		iops += td->io_blocks[DDIR_WRITE];
+		bytes += td->this_io_bytes[DDIR_WRITE];
+	}
+
 	/*
 	 * if rate blocks is set, sample is running
 	 */
-	if (td->rate_bytes) {
+	if (td->rate_bytes || td->rate_blocks) {
 		spent = mtime_since(&td->lastrate, now);
-		if (spent < td->ratecycle)
+		if (spent < td->o.ratecycle)
 			return 0;
 
-		rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
-		if (rate < td->ratemin) {
-			fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
-			return 1;
+		if (td->o.rate) {
+			/*
+			 * check bandwidth specified rate
+			 */
+			if (bytes < td->rate_bytes) {
+				log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
+				return 1;
+			} else {
+				rate = (bytes - td->rate_bytes) / spent;
+				if (rate < td->o.ratemin || bytes < td->rate_bytes) {
+					log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
+					return 1;
+				}
+			}
+		} else {
+			/*
+			 * checks iops specified rate
+			 */
+			if (iops < td->o.rate_iops) {
+				log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
+				return 1;
+			} else {
+				rate = (iops - td->rate_blocks) / spent;
+				if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
+					log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
+				}
+			}
 		}
 	}
 
-	td->rate_bytes = td->this_io_bytes[ddir];
+	td->rate_bytes = bytes;
+	td->rate_blocks = iops;
 	memcpy(&td->lastrate, now, sizeof(*now));
 	return 0;
 }
 
 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 {
-	if (!td->timeout)
+	if (!td->o.timeout)
 		return 0;
-	if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
+	if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 		return 1;
 
 	return 0;
 }
 
-static inline void td_set_runstate(struct thread_data *td, int runstate)
-{
-	td->runstate = runstate;
-}
-
-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.
  */
 static void cleanup_pending_aio(struct thread_data *td)
 {
-	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 	struct list_head *entry, *n;
-	struct io_completion_data icd;
 	struct io_u *io_u;
 	int r;
 
 	/*
 	 * get immediately available events, if any
 	 */
-	r = td_io_getevents(td, 0, td->cur_depth, &ts);
-	if (r > 0) {
-		icd.nr = r;
-		ios_completed(td, &icd);
-	}
+	r = io_u_queued_complete(td, 0);
+	if (r < 0)
+		return;
 
 	/*
 	 * now cancel remaining active events
@@ -179,19 +206,24 @@ 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) {
-		r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
-		if (r > 0) {
-			icd.nr = r;
-			ios_completed(td, &icd);
-		}
-	}
+	if (td->cur_depth)
+		r = io_u_queued_complete(td, td->cur_depth);
 }
 
 /*
@@ -201,7 +233,6 @@ static void cleanup_pending_aio(struct thread_data *td)
 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 {
 	struct io_u *io_u = __get_io_u(td);
-	struct io_completion_data icd;
 	int ret;
 
 	if (!io_u)
@@ -215,54 +246,64 @@ 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) {
-		td_verror(td, io_u->error);
-		put_io_u(td, io_u);
-		return 1;
-	}
-
-	ret = td_io_getevents(td, 1, td->cur_depth, NULL);
 	if (ret < 0) {
-		td_verror(td, ret);
+		td_verror(td, io_u->error, "td_io_queue");
+		put_io_u(td, io_u);
 		return 1;
-	}
+	} else if (ret == FIO_Q_QUEUED) {
+		if (io_u_queued_complete(td, 1) < 0)
+			return 1;
+	} else if (ret == FIO_Q_COMPLETED) {
+		if (io_u->error) {
+			td_verror(td, io_u->error, "td_io_queue");
+			return 1;
+		}
 
-	icd.nr = ret;
-	ios_completed(td, &icd);
-	if (icd.error) {
-		td_verror(td, icd.error);
-		return 1;
+		if (io_u_sync_complete(td, io_u) < 0)
+			return 1;
+	} else if (ret == FIO_Q_BUSY) {
+		if (td_io_commit(td))
+			return 1;
+		goto requeue;
 	}
 
 	return 0;
 }
 
 /*
- * The main verify engine. Runs over the writes we previusly submitted,
+ * The main verify engine. Runs over the writes we previously submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
  */
 static void do_verify(struct thread_data *td)
 {
-	struct io_u *io_u, *v_io_u = NULL;
-	struct io_completion_data icd;
 	struct fio_file *f;
-	int ret, i;
+	struct io_u *io_u;
+	int ret, min_events;
+	unsigned int i;
 
 	/*
 	 * sync io first and invalidate cache, to make sure we really
 	 * read from disk.
 	 */
 	for_each_file(td, f, i) {
-		fio_io_sync(td, f);
-		file_invalidate_cache(td, f);
+		if (!(f->flags & FIO_FILE_OPEN))
+			continue;
+		if (fio_io_sync(td, f))
+			break;
+		if (file_invalidate_cache(td, f))
+			break;
 	}
 
+	if (td->error)
+		return;
+
 	td_set_runstate(td, TD_VERIFYING);
 
-	do {
-		if (td->terminate)
-			break;
+	io_u = NULL;
+	while (!td->terminate) {
+		int ret2;
 
 		io_u = __get_io_u(td);
 		if (!io_u)
@@ -278,195 +319,243 @@ static void do_verify(struct thread_data *td)
 			break;
 		}
 
-		f = get_next_file(td);
-		if (!f)
-			break;
-
-		io_u->file = f;
-
 		if (td_io_prep(td, io_u)) {
 			put_io_u(td, io_u);
 			break;
 		}
 
+		io_u->end_io = verify_io_u;
+
 		ret = td_io_queue(td, io_u);
-		if (ret) {
-			td_verror(td, io_u->error);
-			put_io_u(td, io_u);
+		switch (ret) {
+		case FIO_Q_COMPLETED:
+			if (io_u->error)
+				ret = -io_u->error;
+			else if (io_u->resid) {
+				int bytes = io_u->xfer_buflen - io_u->resid;
+
+				/*
+				 * zero read, fail
+				 */
+				if (!bytes) {
+					td_verror(td, ENODATA, "full resid");
+					put_io_u(td, io_u);
+					break;
+				}
+				io_u->xfer_buflen = io_u->resid;
+				io_u->xfer_buf += bytes;
+				requeue_io_u(td, &io_u);
+			} else {
+				ret = io_u_sync_complete(td, io_u);
+				if (ret < 0)
+					break;
+			}
+			continue;
+		case FIO_Q_QUEUED:
 			break;
-		}
-
-		/*
-		 * we have one pending to verify, do that while
-		 * we are doing io on the next one
-		 */
-		if (do_io_u_verify(td, &v_io_u))
+		case FIO_Q_BUSY:
+			requeue_io_u(td, &io_u);
+			ret2 = td_io_commit(td);
+			if (ret2 < 0)
+				ret = ret2;
 			break;
-
-		ret = td_io_getevents(td, 1, 1, NULL);
-		if (ret != 1) {
-			if (ret < 0)
-				td_verror(td, ret);
+		default:
+			assert(ret < 0);
+			td_verror(td, -ret, "td_io_queue");
 			break;
 		}
 
-		v_io_u = td->io_ops->event(td, 0);
-		icd.nr = 1;
-		icd.error = 0;
-		fio_gettime(&icd.time, NULL);
-		io_completed(td, v_io_u, &icd);
-
-		if (icd.error) {
-			td_verror(td, icd.error);
-			put_io_u(td, v_io_u);
-			v_io_u = NULL;
+		if (ret < 0 || td->error)
 			break;
+
+		/*
+		 * if we can queue more, do so. but check if there are
+		 * completed io_u's first.
+		 */
+		min_events = 0;
+		if (queue_full(td) || ret == FIO_Q_BUSY) {
+			min_events = 1;
+
+			if (td->cur_depth > td->o.iodepth_low)
+				min_events = td->cur_depth - td->o.iodepth_low;
 		}
 
 		/*
-		 * if we can't submit more io, we need to verify now
+		 * Reap required number of io units, if any, and do the
+		 * verification on them through the callback handler
 		 */
-		if (queue_full(td) && do_io_u_verify(td, &v_io_u))
+		if (io_u_queued_complete(td, min_events) < 0)
 			break;
+	}
 
-	} while (1);
-
-	do_io_u_verify(td, &v_io_u);
+	if (!td->error) {
+		min_events = td->cur_depth;
 
-	if (td->cur_depth)
+		if (min_events)
+			ret = io_u_queued_complete(td, min_events);
+	} else
 		cleanup_pending_aio(td);
 
 	td_set_runstate(td, TD_RUNNING);
 }
 
-/*
- * Not really an io thread, all it does is burn CPU cycles in the specified
- * manner.
- */
-static void do_cpuio(struct thread_data *td)
-{
-	struct timeval e;
-	int split = 100 / td->cpuload;
-	int i = 0;
-
-	while (!td->terminate) {
-		fio_gettime(&e, NULL);
-
-		if (runtime_exceeded(td, &e))
-			break;
-
-		if (!(i % split))
-			__usec_sleep(10000);
-		else
-			usec_sleep(td, 10000);
-
-		i++;
-	}
-}
-
 /*
  * Main IO worker function. It retrieves io_u's to process and queues
  * and reaps them, checking for rate and errors along the way.
  */
 static void do_io(struct thread_data *td)
 {
-	struct io_completion_data icd;
 	struct timeval s;
 	unsigned long usec;
-	struct fio_file *f;
-	int i, ret = 0;
+	unsigned int i;
+	int ret = 0;
 
 	td_set_runstate(td, TD_RUNNING);
 
-	while (td->this_io_bytes[td->ddir] < td->io_size) {
-		struct timespec *timeout;
+	while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
+		struct timeval comp_time;
+		long bytes_done = 0;
 		int min_evts = 0;
 		struct io_u *io_u;
+		int ret2;
 
 		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;
 
 		memcpy(&s, &io_u->start_time, sizeof(s));
 
-		ret = td_io_queue(td, io_u);
-		if (ret) {
-			td_verror(td, io_u->error);
+		if (runtime_exceeded(td, &s)) {
 			put_io_u(td, io_u);
 			break;
 		}
 
-		add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
-
-		if (td->cur_depth < td->iodepth) {
-			struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
+		ret = td_io_queue(td, io_u);
+		switch (ret) {
+		case FIO_Q_COMPLETED:
+			if (io_u->error)
+				ret = -io_u->error;
+			else if (io_u->resid) {
+				int bytes = io_u->xfer_buflen - io_u->resid;
+
+				/*
+				 * zero read, fail
+				 */
+				if (!bytes) {
+					td_verror(td, ENODATA, "full resid");
+					put_io_u(td, io_u);
+					break;
+				}
 
-			timeout = &ts;
-			min_evts = 0;
-		} else {
-			timeout = NULL;
-			min_evts = 1;
+				io_u->xfer_buflen = io_u->resid;
+				io_u->xfer_buf += bytes;
+				requeue_io_u(td, &io_u);
+			} else {
+				fio_gettime(&comp_time, NULL);
+				bytes_done = io_u_sync_complete(td, io_u);
+				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);
+			ret2 = td_io_commit(td);
+			if (ret2 < 0)
+				ret = ret2;
+			break;
+		default:
+			assert(ret < 0);
+			put_io_u(td, io_u);
+			break;
 		}
 
-		ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
-		if (ret < 0) {
-			td_verror(td, ret);
+		if (ret < 0 || td->error)
 			break;
-		} else if (!ret)
-			continue;
 
-		icd.nr = ret;
-		ios_completed(td, &icd);
-		if (icd.error) {
-			td_verror(td, icd.error);
-			break;
+		/*
+		 * See if we need to complete some commands
+		 */
+		if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
+			min_evts = 0;
+			if (queue_full(td) || ret == FIO_Q_BUSY) {
+				min_evts = 1;
+
+				if (td->cur_depth > td->o.iodepth_low)
+					min_evts = td->cur_depth - td->o.iodepth_low;
+			}
+
+			fio_gettime(&comp_time, NULL);
+			bytes_done = io_u_queued_complete(td, min_evts);
+			if (bytes_done < 0)
+				break;
 		}
 
+		if (!bytes_done)
+			continue;
+
 		/*
 		 * the rate is batched for now, it should work for batches
 		 * of completions except the very first one which may look
 		 * a little bursty
 		 */
-		usec = utime_since(&s, &icd.time);
+		usec = utime_since(&s, &comp_time);
 
-		rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
+		rate_throttle(td, usec, bytes_done);
 
-		if (check_min_rate(td, &icd.time)) {
+		if (check_min_rate(td, &comp_time)) {
 			if (exitall_on_terminate)
 				terminate_threads(td->groupid);
-			td_verror(td, ENOMEM);
+			td_verror(td, ENODATA, "check_min_rate");
 			break;
 		}
 
-		if (runtime_exceeded(td, &icd.time))
-			break;
-
-		if (td->thinktime) {
+		if (td->o.thinktime) {
 			unsigned long long b;
 
 			b = td->io_blocks[0] + td->io_blocks[1];
-			if (!(td->thinktime_blocks % b))
-				usec_sleep(td, td->thinktime);
+			if (!(b % td->o.thinktime_blocks)) {
+				int left;
+
+				if (td->o.thinktime_spin)
+					__usec_sleep(td->o.thinktime_spin);
+
+				left = td->o.thinktime - td->o.thinktime_spin;
+				if (left)
+					usec_sleep(td, left);
+			}
 		}
 	}
 
 	if (!td->error) {
-		if (td->cur_depth)
-			cleanup_pending_aio(td);
+		struct fio_file *f;
 
-		if (should_fsync(td) && td->end_fsync) {
+		i = td->cur_depth;
+		if (i)
+			ret = io_u_queued_complete(td, i);
+
+		if (should_fsync(td) && td->o.end_fsync) {
 			td_set_runstate(td, TD_FSYNCING);
-			for_each_file(td, f, i)
+
+			for_each_file(td, f, i) {
+				if (!(f->flags & FIO_FILE_OPEN))
+					continue;
 				fio_io_sync(td, f);
+			}
 		}
-	}
+	} else
+		cleanup_pending_aio(td);
 }
 
 static void cleanup_io_u(struct thread_data *td)
@@ -499,26 +588,31 @@ static void fill_rand_buf(struct io_u *io_u, int max_bs)
 
 static int init_io_u(struct thread_data *td)
 {
+	unsigned long long buf_size;
 	struct io_u *io_u;
 	unsigned int max_bs;
 	int i, max_units;
 	char *p;
 
-	if (td->io_ops->flags & FIO_CPUIO)
-		return 0;
-
 	if (td->io_ops->flags & FIO_SYNCIO)
 		max_units = 1;
 	else
-		max_units = td->iodepth;
+		max_units = td->o.iodepth;
+
+	max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
+	buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
+	buf_size += page_mask;
+	if (buf_size != (size_t) buf_size) {
+		log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
+		return 1;
+	}
 
-	max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
-	td->orig_buffer_size = max_bs * max_units;
+	td->orig_buffer_size = buf_size;
 
-	if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
-		td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
-	else
-		td->orig_buffer_size += MASK;
+	if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
+		td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
+	else if (td->orig_buffer_size & page_mask)
+		td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
 
 	if (allocate_io_mem(td))
 		return 1;
@@ -534,9 +628,12 @@ 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;
 }
 
@@ -546,23 +643,23 @@ static int switch_ioscheduler(struct thread_data *td)
 	FILE *f;
 	int ret;
 
-	if (td->io_ops->flags & FIO_CPUIO)
+	if (td->io_ops->flags & FIO_DISKLESSIO)
 		return 0;
 
 	sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 
 	f = fopen(tmp, "r+");
 	if (!f) {
-		td_verror(td, errno);
+		td_verror(td, errno, "fopen");
 		return 1;
 	}
 
 	/*
 	 * Set io scheduler.
 	 */
-	ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
+	ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 	if (ferror(f) || ret != 1) {
-		td_verror(td, errno);
+		td_verror(td, errno, "fwrite");
 		fclose(f);
 		return 1;
 	}
@@ -574,15 +671,15 @@ static int switch_ioscheduler(struct thread_data *td)
 	 */
 	ret = fread(tmp, 1, sizeof(tmp), f);
 	if (ferror(f) || ret < 0) {
-		td_verror(td, errno);
+		td_verror(td, errno, "fread");
 		fclose(f);
 		return 1;
 	}
 
-	sprintf(tmp2, "[%s]", td->ioscheduler);
+	sprintf(tmp2, "[%s]", td->o.ioscheduler);
 	if (!strstr(tmp, tmp2)) {
-		log_err("fio: io scheduler %s not found\n", td->ioscheduler);
-		td_verror(td, EINVAL);
+		log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
+		td_verror(td, EINVAL, "iosched_switch");
 		fclose(f);
 		return 1;
 	}
@@ -591,23 +688,32 @@ static int switch_ioscheduler(struct thread_data *td)
 	return 0;
 }
 
-static void clear_io_state(struct thread_data *td)
+static int clear_io_state(struct thread_data *td)
 {
 	struct fio_file *f;
-	int i;
+	unsigned int i;
+	int ret;
 
-	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->rate_bytes = 0;
+	td->rate_blocks = 0;
+	td->rw_end_set[0] = td->rw_end_set[1] = 0;
 
-	for_each_file(td, f, i) {
-		f->last_pos = 0;
-		if (td->io_ops->flags & FIO_SYNCIO)
-			lseek(f->fd, SEEK_SET, 0);
+	td->last_was_sync = 0;
+
+	for_each_file(td, f, i)
+		td_io_close_file(td, f);
 
-		if (f->file_map)
-			memset(f->file_map, 0, f->num_maps * sizeof(long));
+	ret = 0;
+	for_each_file(td, f, i) {
+		ret = td_io_open_file(td, f);
+		if (ret)
+			break;
 	}
+
+	return ret;
 }
 
 /*
@@ -618,91 +724,120 @@ static void *thread_main(void *data)
 {
 	unsigned long long runtime[2];
 	struct thread_data *td = data;
+	unsigned long elapsed;
+	int clear_state;
 
-	if (!td->use_thread)
+	if (!td->o.use_thread)
 		setsid();
 
 	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_u_requeues);
 	INIT_LIST_HEAD(&td->io_log_list);
 
 	if (init_io_u(td))
-		goto err;
+		goto err_sem;
 
 	if (fio_setaffinity(td) == -1) {
-		td_verror(td, errno);
-		goto err;
+		td_verror(td, errno, "cpu_set_affinity");
+		goto err_sem;
 	}
 
-	if (td_io_init(td))
-		goto err;
-
 	if (init_iolog(td))
-		goto err;
+		goto err_sem;
 
 	if (td->ioprio) {
 		if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
-			td_verror(td, errno);
-			goto err;
+			td_verror(td, errno, "ioprio_set");
+			goto err_sem;
 		}
 	}
 
-	if (nice(td->nice) == -1) {
-		td_verror(td, errno);
-		goto err;
+	if (nice(td->o.nice) == -1) {
+		td_verror(td, errno, "nice");
+		goto err_sem;
 	}
 
-	if (init_random_state(td))
-		goto err;
-
-	if (td->ioscheduler && switch_ioscheduler(td))
-		goto err;
+	if (td->o.ioscheduler && switch_ioscheduler(td))
+		goto err_sem;
 
 	td_set_runstate(td, TD_INITIALIZED);
-	fio_sem_up(&startup_sem);
-	fio_sem_down(&td->mutex);
+	fio_sem_up(startup_sem);
+	fio_sem_down(td->mutex);
+
+	/*
+	 * the ->mutex semaphore is now no longer used, close it to avoid
+	 * eating a file descriptor
+	 */
+	fio_sem_remove(td->mutex);
+
+	if (!td->o.create_serialize && setup_files(td))
+		goto err;
 
-	if (!td->create_serialize && setup_files(td))
+	if (td_io_init(td))
 		goto err;
+
 	if (open_files(td))
 		goto err;
 
-	if (td->exec_prerun)
-		system(td->exec_prerun);
+	if (init_random_map(td))
+		goto err;
+
+	if (td->o.exec_prerun) {
+		if (system(td->o.exec_prerun) < 0)
+			goto err;
+	}
 
 	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--) {
+	clear_state = 0;
+	while (td->o.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));
+		if (td->o.ratemin)
+			memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
+
+		if (clear_state && clear_io_state(td))
+			break;
 
-		clear_io_state(td);
 		prune_io_piece_log(td);
 
-		if (td->io_ops->flags & FIO_CPUIO)
-			do_cpuio(td);
-		else
-			do_io(td);
+		do_io(td);
+
+		clear_state = 1;
 
-		runtime[td->ddir] += utime_since_now(&td->start);
-		if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
-			runtime[td->ddir ^ 1] = runtime[td->ddir];
+		if (td_read(td) && td->io_bytes[DDIR_READ]) {
+			if (td->rw_end_set[DDIR_READ])
+				elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
+			else
+				elapsed = utime_since_now(&td->start);
+
+			runtime[DDIR_READ] += elapsed;
+		}
+		if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
+			if (td->rw_end_set[DDIR_WRITE])
+				elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
+			else
+				elapsed = utime_since_now(&td->start);
 
+			runtime[DDIR_WRITE] += elapsed;
+		}
+		
 		if (td->error || td->terminate)
 			break;
 
-		if (td->verify == VERIFY_NONE)
+		if (td->o.verify == VERIFY_NONE)
 			continue;
 
-		clear_io_state(td);
+		if (clear_io_state(td))
+			break;
+
 		fio_gettime(&td->start, NULL);
 
 		do_verify(td);
@@ -714,52 +849,63 @@ static void *thread_main(void *data)
 	}
 
 	update_rusage_stat(td);
-	fio_gettime(&td->end_time, NULL);
-	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->write_iolog_file)
+	td->ts.runtime[0] = runtime[0] / 1000;
+	td->ts.runtime[1] = runtime[1] / 1000;
+	td->ts.total_run_time = mtime_since_now(&td->epoch);
+	td->ts.io_bytes[0] = td->io_bytes[0];
+	td->ts.io_bytes[1] = td->io_bytes[1];
+
+	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->o.write_iolog_file)
 		write_iolog_close(td);
-	if (td->exec_postrun)
-		system(td->exec_postrun);
+	if (td->o.exec_postrun) {
+		if (system(td->o.exec_postrun) < 0)
+			log_err("fio: postrun %s failed\n", td->o.exec_postrun);
+	}
 
 	if (exitall_on_terminate)
 		terminate_threads(td->groupid);
 
 err:
+	if (td->error)
+		printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 	close_files(td);
 	close_ioengine(td);
 	cleanup_io_u(td);
+	options_mem_free(td);
 	td_set_runstate(td, TD_EXITED);
-	return NULL;
-
+	return (void *) (unsigned long) td->error;
+err_sem:
+	fio_sem_up(startup_sem);
+	goto err;
 }
 
 /*
  * We cannot pass the td data into a forked process, so attach the td and
  * pass it to the thread worker.
  */
-static void *fork_main(int shmid, int offset)
+static int fork_main(int shmid, int offset)
 {
 	struct thread_data *td;
-	void *data;
+	void *data, *ret;
 
 	data = shmat(shmid, NULL, 0);
 	if (data == (void *) -1) {
+		int __err = errno;
+
 		perror("shmat");
-		return NULL;
+		return __err;
 	}
 
 	td = data + offset * sizeof(struct thread_data);
-	thread_main(td);
+	ret = thread_main(td);
 	shmdt(data);
-	return NULL;
+	return (int) (unsigned long) ret;
 }
 
 /*
@@ -768,40 +914,83 @@ static void *fork_main(int shmid, int offset)
 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 {
 	struct thread_data *td;
-	int i, cputhreads, pending;
+	int i, cputhreads, pending, status, ret;
 
 	/*
 	 * reap exited threads (TD_EXITED -> TD_REAPED)
 	 */
 	pending = cputhreads = 0;
 	for_each_td(td, i) {
+		int flags = 0;
+
 		/*
 		 * ->io_ops is NULL for a thread that has closed its
 		 * io engine
 		 */
-		if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
+		if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 			cputhreads++;
 
-		if (td->runstate != TD_EXITED) {
-			if (td->runstate < TD_RUNNING)
-				pending++;
-
+		if (!td->pid || td->runstate == TD_REAPED)
+			continue;
+		if (td->o.use_thread) {
+			if (td->runstate == TD_EXITED) {
+				td_set_runstate(td, TD_REAPED);
+				goto reaped;
+			}
 			continue;
 		}
 
-		td_set_runstate(td, TD_REAPED);
+		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) {
+			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);
+
+				if (sig != SIGQUIT)
+					log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
+				td_set_runstate(td, TD_REAPED);
+				goto reaped;
+			}
+			if (WIFEXITED(status)) {
+				if (WEXITSTATUS(status) && !td->error)
+					td->error = WEXITSTATUS(status);
+
+				td_set_runstate(td, TD_REAPED);
+				goto reaped;
+			}
+		}
 
-		if (td->use_thread) {
+		/*
+		 * thread is not dead, continue
+		 */
+		continue;
+reaped:
+		if (td->o.use_thread) {
 			long ret;
 
 			if (pthread_join(td->thread, (void *) &ret))
-				perror("thread_join");
-		} else
-			waitpid(td->pid, NULL, 0);
+				perror("pthread_join");
+		}
 
 		(*nr_running)--;
-		(*m_rate) -= td->ratemin;
-		(*t_rate) -= td->rate;
+		(*m_rate) -= td->o.ratemin;
+		(*t_rate) -= td->o.rate;
+
+		if (td->error)
+			exit_value++;
 	}
 
 	if (*nr_running == cputhreads && !pending)
@@ -821,7 +1010,15 @@ static void run_threads(void)
 		return;
 
 	if (!terse_output) {
-		printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
+		printf("Starting ");
+		if (nr_thread)
+			printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
+		if (nr_process) {
+			if (nr_thread)
+				printf(" and ");
+			printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
+		}
+		printf("\n");
 		fflush(stdout);
 	}
 
@@ -836,10 +1033,10 @@ static void run_threads(void)
 	for_each_td(td, i) {
 		print_status_init(td->thread_number - 1);
 
-		init_disk_util(td);
-
-		if (!td->create_serialize)
+		if (!td->o.create_serialize) {
+			init_disk_util(td);
 			continue;
+		}
 
 		/*
 		 * do file setup here so it happens sequentially,
@@ -847,11 +1044,18 @@ static void run_threads(void)
 		 * client data interspersed on disk
 		 */
 		if (setup_files(td)) {
+			exit_value++;
+			if (td->error)
+				log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 			td_set_runstate(td, TD_REAPED);
 			todo--;
 		}
+
+		init_disk_util(td);
 	}
 
+	set_genesis_time();
+
 	while (todo) {
 		struct thread_data *map[MAX_JOBS];
 		struct timeval this_start;
@@ -873,14 +1077,14 @@ static void run_threads(void)
 				continue;
 			}
 
-			if (td->start_delay) {
+			if (td->o.start_delay) {
 				spent = mtime_since_genesis();
 
-				if (td->start_delay * 1000 > spent)
+				if (td->o.start_delay * 1000 > spent)
 					continue;
 			}
 
-			if (td->stonewall && (nr_started || nr_running))
+			if (td->o.stonewall && (nr_started || nr_running))
 				break;
 
 			/*
@@ -889,22 +1093,22 @@ static void run_threads(void)
 			 */
 			td_set_runstate(td, TD_CREATED);
 			map[this_jobs++] = td;
-			fio_sem_init(&startup_sem, 1);
 			nr_started++;
 
-			if (td->use_thread) {
+			if (td->o.use_thread) {
 				if (pthread_create(&td->thread, NULL, thread_main, td)) {
 					perror("thread_create");
 					nr_started--;
+					break;
 				}
 			} else {
-				if (fork())
-					fio_sem_down(&startup_sem);
-				else {
-					fork_main(shm_id, i);
-					exit(0);
+				if (!fork()) {
+					int ret = fork_main(shm_id, i);
+
+					exit(ret);
 				}
 			}
+			fio_sem_down(startup_sem);
 		}
 
 		/*
@@ -913,7 +1117,7 @@ static void run_threads(void)
 		 */
 		fio_gettime(&this_start, NULL);
 		left = this_jobs;
-		while (left) {
+		while (left && !fio_abort) {
 			if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 				break;
 
@@ -956,10 +1160,10 @@ static void run_threads(void)
 			td_set_runstate(td, TD_RUNNING);
 			nr_running++;
 			nr_started--;
-			m_rate += td->ratemin;
-			t_rate += td->rate;
+			m_rate += td->o.ratemin;
+			t_rate += td->o.rate;
 			todo--;
-			fio_sem_up(&td->mutex);
+			fio_sem_up(td->mutex);
 		}
 
 		reap_threads(&nr_running, &t_rate, &m_rate);
@@ -979,6 +1183,15 @@ static void run_threads(void)
 
 int main(int argc, char *argv[])
 {
+	long ps;
+
+	/*
+	 * We need locale for number printing, if it isn't set then just
+	 * go with the US format.
+	 */
+	if (!getenv("LC_NUMERIC"))
+		setlocale(LC_NUMERIC, "en_US");
+
 	if (parse_options(argc, argv))
 		return 1;
 
@@ -987,10 +1200,36 @@ int main(int argc, char *argv[])
 		return 1;
 	}
 
+	ps = sysconf(_SC_PAGESIZE);
+	if (ps < 0) {
+		log_err("Failed to get page size\n");
+		return 1;
+	}
+
+	page_size = ps;
+	page_mask = ps - 1;
+
+	if (write_bw_log) {
+		setup_log(&agg_io_log[DDIR_READ]);
+		setup_log(&agg_io_log[DDIR_WRITE]);
+	}
+
+	startup_sem = fio_sem_init(0);
+
+	set_genesis_time();
+
 	disk_util_timer_arm();
 
 	run_threads();
-	show_run_stats();
 
-	return 0;
+	if (!fio_abort) {
+		show_run_stats();
+		if (write_bw_log) {
+			__finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
+			__finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
+		}
+	}
+
+	fio_sem_remove(startup_sem);
+	return exit_value;
 }