X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=backend.c;h=c208b227846895c63d432c62531689a074dda66c;hp=1dafe68d2e36fa97f85f6f7950b86a7728d656de;hb=17ecadc3fababaace961051765f140d3332d362d;hpb=55312f9f5572611037b6254912460f64173df574

diff --git a/backend.c b/backend.c
index 1dafe68d..c208b227 100644
--- a/backend.c
+++ b/backend.c
@@ -35,6 +35,7 @@
 #include <sys/wait.h>
 #include <sys/ipc.h>
 #include <sys/mman.h>
+#include <math.h>
 
 #include "fio.h"
 #ifndef FIO_NO_HAVE_SHM_H
@@ -48,12 +49,14 @@
 #include "cgroup.h"
 #include "profile.h"
 #include "lib/rand.h"
-#include "memalign.h"
+#include "lib/memalign.h"
 #include "server.h"
 #include "lib/getrusage.h"
 #include "idletime.h"
 #include "err.h"
 #include "lib/tp.h"
+#include "workqueue.h"
+#include "lib/mountcheck.h"
 
 static pthread_t helper_thread;
 static pthread_mutex_t helper_lock;
@@ -98,7 +101,7 @@ static void sig_int(int sig)
 	}
 }
 
-static void sig_show_status(int sig)
+void sig_show_status(int sig)
 {
 	show_running_run_stats();
 }
@@ -141,8 +144,8 @@ static void set_sig_handlers(void)
 /*
  * Check if we are above the minimum rate given.
  */
-static int __check_min_rate(struct thread_data *td, struct timeval *now,
-			    enum fio_ddir ddir)
+static bool __check_min_rate(struct thread_data *td, struct timeval *now,
+			     enum fio_ddir ddir)
 {
 	unsigned long long bytes = 0;
 	unsigned long iops = 0;
@@ -155,13 +158,13 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
 	assert(ddir_rw(ddir));
 
 	if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
-		return 0;
+		return false;
 
 	/*
 	 * allow a 2 second settle period in the beginning
 	 */
 	if (mtime_since(&td->start, now) < 2000)
-		return 0;
+		return false;
 
 	iops += td->this_io_blocks[ddir];
 	bytes += td->this_io_bytes[ddir];
@@ -175,16 +178,16 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
 	if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 		spent = mtime_since(&td->lastrate[ddir], now);
 		if (spent < td->o.ratecycle)
-			return 0;
+			return false;
 
-		if (td->o.rate[ddir]) {
+		if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 			/*
 			 * check bandwidth specified rate
 			 */
 			if (bytes < td->rate_bytes[ddir]) {
 				log_err("%s: min rate %u not met\n", td->o.name,
 								ratemin);
-				return 1;
+				return true;
 			} else {
 				if (spent)
 					rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
@@ -196,7 +199,7 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
 					log_err("%s: min rate %u not met, got"
 						" %luKB/sec\n", td->o.name,
 							ratemin, rate);
-					return 1;
+					return true;
 				}
 			}
 		} else {
@@ -206,7 +209,7 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
 			if (iops < rate_iops) {
 				log_err("%s: min iops rate %u not met\n",
 						td->o.name, rate_iops);
-				return 1;
+				return true;
 			} else {
 				if (spent)
 					rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
@@ -218,6 +221,7 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
 					log_err("%s: min iops rate %u not met,"
 						" got %lu\n", td->o.name,
 							rate_iops_min, rate);
+					return true;
 				}
 			}
 		}
@@ -226,12 +230,12 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
 	td->rate_bytes[ddir] = bytes;
 	td->rate_blocks[ddir] = iops;
 	memcpy(&td->lastrate[ddir], now, sizeof(*now));
-	return 0;
+	return false;
 }
 
-static int check_min_rate(struct thread_data *td, struct timeval *now)
+static bool check_min_rate(struct thread_data *td, struct timeval *now)
 {
-	int ret = 0;
+	bool ret = false;
 
 	if (td->bytes_done[DDIR_READ])
 		ret |= __check_min_rate(td, now, DDIR_READ);
@@ -282,20 +286,20 @@ static void cleanup_pending_aio(struct thread_data *td)
  * Helper to handle the final sync of a file. Works just like the normal
  * io path, just does everything sync.
  */
-static int fio_io_sync(struct thread_data *td, struct fio_file *f)
+static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 {
 	struct io_u *io_u = __get_io_u(td);
 	int ret;
 
 	if (!io_u)
-		return 1;
+		return true;
 
 	io_u->ddir = DDIR_SYNC;
 	io_u->file = f;
 
 	if (td_io_prep(td, io_u)) {
 		put_io_u(td, io_u);
-		return 1;
+		return true;
 	}
 
 requeue:
@@ -303,25 +307,25 @@ requeue:
 	if (ret < 0) {
 		td_verror(td, io_u->error, "td_io_queue");
 		put_io_u(td, io_u);
-		return 1;
+		return true;
 	} else if (ret == FIO_Q_QUEUED) {
 		if (io_u_queued_complete(td, 1) < 0)
-			return 1;
+			return true;
 	} else if (ret == FIO_Q_COMPLETED) {
 		if (io_u->error) {
 			td_verror(td, io_u->error, "td_io_queue");
-			return 1;
+			return true;
 		}
 
 		if (io_u_sync_complete(td, io_u) < 0)
-			return 1;
+			return true;
 	} else if (ret == FIO_Q_BUSY) {
 		if (td_io_commit(td))
-			return 1;
+			return true;
 		goto requeue;
 	}
 
-	return 0;
+	return false;
 }
 
 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
@@ -350,20 +354,37 @@ static inline void update_tv_cache(struct thread_data *td)
 		__update_tv_cache(td);
 }
 
-static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
+static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
 {
 	if (in_ramp_time(td))
-		return 0;
+		return false;
 	if (!td->o.timeout)
-		return 0;
+		return false;
 	if (utime_since(&td->epoch, t) >= td->o.timeout)
-		return 1;
+		return true;
 
-	return 0;
+	return false;
 }
 
-static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
-			       int *retptr)
+/*
+ * We need to update the runtime consistently in ms, but keep a running
+ * tally of the current elapsed time in microseconds for sub millisecond
+ * updates.
+ */
+static inline void update_runtime(struct thread_data *td,
+				  unsigned long long *elapsed_us,
+				  const enum fio_ddir ddir)
+{
+	if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
+		return;
+
+	td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
+	elapsed_us[ddir] += utime_since_now(&td->start);
+	td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
+}
+
+static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
+				int *retptr)
 {
 	int ret = *retptr;
 
@@ -376,7 +397,7 @@ static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 
 		eb = td_error_type(ddir, err);
 		if (!(td->o.continue_on_error & (1 << eb)))
-			return 1;
+			return true;
 
 		if (td_non_fatal_error(td, eb, err)) {
 		        /*
@@ -386,7 +407,7 @@ static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 			update_error_count(td, err);
 			td_clear_error(td);
 			*retptr = 0;
-			return 0;
+			return false;
 		} else if (td->o.fill_device && err == ENOSPC) {
 			/*
 			 * We expect to hit this error if
@@ -394,18 +415,18 @@ static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 			 */
 			td_clear_error(td);
 			fio_mark_td_terminate(td);
-			return 1;
+			return true;
 		} else {
 			/*
 			 * Stop the I/O in case of a fatal
 			 * error.
 			 */
 			update_error_count(td, err);
-			return 1;
+			return true;
 		}
 	}
 
-	return 0;
+	return false;
 }
 
 static void check_update_rusage(struct thread_data *td)
@@ -426,8 +447,8 @@ static int wait_for_completions(struct thread_data *td, struct timeval *time)
 	/*
 	 * if the queue is full, we MUST reap at least 1 event
 	 */
-	min_evts = min(td->o.iodepth_batch_complete, td->cur_depth);
-	if (full && !min_evts)
+	min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
+	if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 		min_evts = 1;
 
 	if (time && (__should_check_rate(td, DDIR_READ) ||
@@ -444,6 +465,99 @@ static int wait_for_completions(struct thread_data *td, struct timeval *time)
 	return ret;
 }
 
+int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
+		   enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
+		   struct timeval *comp_time)
+{
+	int ret2;
+
+	switch (*ret) {
+	case FIO_Q_COMPLETED:
+		if (io_u->error) {
+			*ret = -io_u->error;
+			clear_io_u(td, io_u);
+		} else if (io_u->resid) {
+			int bytes = io_u->xfer_buflen - io_u->resid;
+			struct fio_file *f = io_u->file;
+
+			if (bytes_issued)
+				*bytes_issued += bytes;
+
+			if (!from_verify)
+				trim_io_piece(td, io_u);
+
+			/*
+			 * zero read, fail
+			 */
+			if (!bytes) {
+				if (!from_verify)
+					unlog_io_piece(td, io_u);
+				td_verror(td, EIO, "full resid");
+				put_io_u(td, io_u);
+				break;
+			}
+
+			io_u->xfer_buflen = io_u->resid;
+			io_u->xfer_buf += bytes;
+			io_u->offset += bytes;
+
+			if (ddir_rw(io_u->ddir))
+				td->ts.short_io_u[io_u->ddir]++;
+
+			f = io_u->file;
+			if (io_u->offset == f->real_file_size)
+				goto sync_done;
+
+			requeue_io_u(td, &io_u);
+		} else {
+sync_done:
+			if (comp_time && (__should_check_rate(td, DDIR_READ) ||
+			    __should_check_rate(td, DDIR_WRITE) ||
+			    __should_check_rate(td, DDIR_TRIM)))
+				fio_gettime(comp_time, NULL);
+
+			*ret = io_u_sync_complete(td, io_u);
+			if (*ret < 0)
+				break;
+		}
+		return 0;
+	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);
+		if (bytes_issued)
+			*bytes_issued += io_u->xfer_buflen;
+		break;
+	case FIO_Q_BUSY:
+		if (!from_verify)
+			unlog_io_piece(td, io_u);
+		requeue_io_u(td, &io_u);
+		ret2 = td_io_commit(td);
+		if (ret2 < 0)
+			*ret = ret2;
+		break;
+	default:
+		assert(*ret < 0);
+		td_verror(td, -(*ret), "td_io_queue");
+		break;
+	}
+
+	if (break_on_this_error(td, ddir, ret))
+		return 1;
+
+	return 0;
+}
+
+static inline bool io_in_polling(struct thread_data *td)
+{
+	return !td->o.iodepth_batch_complete_min &&
+		   !td->o.iodepth_batch_complete_max;
+}
+
 /*
  * The main verify engine. Runs over the writes we previously submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
@@ -480,7 +594,7 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 	io_u = NULL;
 	while (!td->terminate) {
 		enum fio_ddir ddir;
-		int ret2, full;
+		int full;
 
 		update_tv_cache(td);
 		check_update_rusage(td);
@@ -536,7 +650,7 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 					continue;
 				} else if (io_u->ddir == DDIR_TRIM) {
 					io_u->ddir = DDIR_READ;
-					io_u->flags |= IO_U_F_TRIMMED;
+					io_u_set(io_u, IO_U_F_TRIMMED);
 					break;
 				} else if (io_u->ddir == DDIR_WRITE) {
 					io_u->ddir = DDIR_READ;
@@ -566,57 +680,8 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 			fio_gettime(&io_u->start_time, NULL);
 
 		ret = td_io_queue(td, io_u);
-		switch (ret) {
-		case FIO_Q_COMPLETED:
-			if (io_u->error) {
-				ret = -io_u->error;
-				clear_io_u(td, io_u);
-			} else if (io_u->resid) {
-				int bytes = io_u->xfer_buflen - io_u->resid;
 
-				/*
-				 * zero read, fail
-				 */
-				if (!bytes) {
-					td_verror(td, EIO, "full resid");
-					put_io_u(td, io_u);
-					break;
-				}
-
-				io_u->xfer_buflen = io_u->resid;
-				io_u->xfer_buf += bytes;
-				io_u->offset += bytes;
-
-				if (ddir_rw(io_u->ddir))
-					td->ts.short_io_u[io_u->ddir]++;
-
-				f = io_u->file;
-				if (io_u->offset == f->real_file_size)
-					goto sync_done;
-
-				requeue_io_u(td, &io_u);
-			} else {
-sync_done:
-				ret = io_u_sync_complete(td, io_u);
-				if (ret < 0)
-					break;
-			}
-			continue;
-		case FIO_Q_QUEUED:
-			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);
-			td_verror(td, -ret, "td_io_queue");
-			break;
-		}
-
-		if (break_on_this_error(td, ddir, &ret))
+		if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 			break;
 
 		/*
@@ -626,7 +691,7 @@ sync_done:
 		 */
 reap:
 		full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
-		if (full || !td->o.iodepth_batch_complete)
+		if (full || io_in_polling(td))
 			ret = wait_for_completions(td, NULL);
 
 		if (ret < 0)
@@ -648,12 +713,12 @@ reap:
 	dprint(FD_VERIFY, "exiting loop\n");
 }
 
-static unsigned int exceeds_number_ios(struct thread_data *td)
+static bool exceeds_number_ios(struct thread_data *td)
 {
 	unsigned long long number_ios;
 
 	if (!td->o.number_ios)
-		return 0;
+		return false;
 
 	number_ios = ddir_rw_sum(td->io_blocks);
 	number_ios += td->io_u_queued + td->io_u_in_flight;
@@ -661,7 +726,7 @@ static unsigned int exceeds_number_ios(struct thread_data *td)
 	return number_ios >= (td->o.number_ios * td->loops);
 }
 
-static int io_issue_bytes_exceeded(struct thread_data *td)
+static bool io_issue_bytes_exceeded(struct thread_data *td)
 {
 	unsigned long long bytes, limit;
 
@@ -683,7 +748,7 @@ static int io_issue_bytes_exceeded(struct thread_data *td)
 	return bytes >= limit || exceeds_number_ios(td);
 }
 
-static int io_complete_bytes_exceeded(struct thread_data *td)
+static bool io_complete_bytes_exceeded(struct thread_data *td)
 {
 	unsigned long long bytes, limit;
 
@@ -705,6 +770,38 @@ static int io_complete_bytes_exceeded(struct thread_data *td)
 	return bytes >= limit || exceeds_number_ios(td);
 }
 
+/*
+ * used to calculate the next io time for rate control
+ *
+ */
+static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
+{
+	uint64_t secs, remainder, bps, bytes, iops;
+
+	assert(!(td->flags & TD_F_CHILD));
+	bytes = td->rate_io_issue_bytes[ddir];
+	bps = td->rate_bps[ddir];
+
+	if (td->o.rate_process == RATE_PROCESS_POISSON) {
+		uint64_t val;
+		iops = bps / td->o.bs[ddir];
+		val = (int64_t) (1000000 / iops) *
+				-logf(__rand_0_1(&td->poisson_state));
+		if (val) {
+			dprint(FD_RATE, "poisson rate iops=%llu\n",
+					(unsigned long long) 1000000 / val);
+		}
+		td->last_usec += val;
+		return td->last_usec;
+	} else if (bps) {
+		secs = bytes / bps;
+		remainder = bytes % bps;
+		return remainder * 1000000 / bps + secs * 1000000;
+	}
+
+	return 0;
+}
+
 /*
  * Main IO worker function. It retrieves io_u's to process and queues
  * and reaps them, checking for rate and errors along the way.
@@ -750,7 +847,7 @@ static uint64_t do_io(struct thread_data *td)
 		td->o.time_based) {
 		struct timeval comp_time;
 		struct io_u *io_u;
-		int ret2, full;
+		int full;
 		enum fio_ddir ddir;
 
 		check_update_rusage(td);
@@ -771,7 +868,7 @@ static uint64_t do_io(struct thread_data *td)
 		if (flow_threshold_exceeded(td))
 			continue;
 
-		if (bytes_issued >= total_bytes)
+		if (!td->o.time_based && bytes_issued >= total_bytes)
 			break;
 
 		io_u = get_io_u(td);
@@ -830,90 +927,48 @@ static uint64_t do_io(struct thread_data *td)
 		    !td->o.experimental_verify)
 			log_io_piece(td, io_u);
 
-		ret = td_io_queue(td, io_u);
-		switch (ret) {
-		case FIO_Q_COMPLETED:
-			if (io_u->error) {
-				ret = -io_u->error;
-				unlog_io_piece(td, io_u);
-				clear_io_u(td, io_u);
-			} else if (io_u->resid) {
-				int bytes = io_u->xfer_buflen - io_u->resid;
-				struct fio_file *f = io_u->file;
-
-				bytes_issued += bytes;
+		if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
+			const unsigned long blen = io_u->xfer_buflen;
+			const enum fio_ddir ddir = acct_ddir(io_u);
 
-				trim_io_piece(td, io_u);
+			if (td->error)
+				break;
 
-				/*
-				 * zero read, fail
-				 */
-				if (!bytes) {
-					unlog_io_piece(td, io_u);
-					td_verror(td, EIO, "full resid");
-					put_io_u(td, io_u);
-					break;
-				}
+			ret = workqueue_enqueue(&td->io_wq, &io_u->work);
+			if (ret)
+				ret = FIO_Q_QUEUED;
+			else
+				ret = FIO_Q_BUSY;
 
-				io_u->xfer_buflen = io_u->resid;
-				io_u->xfer_buf += bytes;
-				io_u->offset += bytes;
+			if (ret == FIO_Q_QUEUED && ddir_rw(ddir)) {
+				td->io_issues[ddir]++;
+				td->io_issue_bytes[ddir] += blen;
+				td->rate_io_issue_bytes[ddir] += blen;
+			}
 
-				if (ddir_rw(io_u->ddir))
-					td->ts.short_io_u[io_u->ddir]++;
+			if (should_check_rate(td))
+				td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 
-				if (io_u->offset == f->real_file_size)
-					goto sync_done;
+		} else {
+			ret = td_io_queue(td, io_u);
 
-				requeue_io_u(td, &io_u);
-			} else {
-sync_done:
-				if (__should_check_rate(td, DDIR_READ) ||
-				    __should_check_rate(td, DDIR_WRITE) ||
-				    __should_check_rate(td, DDIR_TRIM))
-					fio_gettime(&comp_time, NULL);
+			if (should_check_rate(td))
+				td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
+
+			if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
+				break;
 
-				ret = io_u_sync_complete(td, io_u);
-				if (ret < 0)
-					break;
-				bytes_issued += io_u->xfer_buflen;
-			}
-			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.
+			 * See if we need to complete some commands. Note that
+			 * we can get BUSY even without IO queued, if the
+			 * system is resource starved.
 			 */
-			if (td->io_ops->commit == NULL)
-				io_u_queued(td, io_u);
-			bytes_issued += io_u->xfer_buflen;
-			break;
-		case FIO_Q_BUSY:
-			unlog_io_piece(td, io_u);
-			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;
-		}
-
-		if (break_on_this_error(td, ddir, &ret))
-			break;
-
-		/*
-		 * See if we need to complete some commands. Note that we
-		 * can get BUSY even without IO queued, if the system is
-		 * resource starved.
-		 */
 reap:
-		full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
-		if (full || !td->o.iodepth_batch_complete)
-			ret = wait_for_completions(td, &comp_time);
+			full = queue_full(td) ||
+				(ret == FIO_Q_BUSY && td->cur_depth);
+			if (full || io_in_polling(td))
+				ret = wait_for_completions(td, &comp_time);
+		}
 		if (ret < 0)
 			break;
 		if (!ddir_rw_sum(td->bytes_done) &&
@@ -962,7 +1017,12 @@ reap:
 	if (!td->error) {
 		struct fio_file *f;
 
-		i = td->cur_depth;
+		if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
+			workqueue_flush(&td->io_wq);
+			i = 0;
+		} else
+			i = td->cur_depth;
+
 		if (i) {
 			ret = io_u_queued_complete(td, i);
 			if (td->o.fill_device && td->error == ENOSPC)
@@ -1176,13 +1236,17 @@ static int switch_ioscheduler(struct thread_data *td)
 	/*
 	 * Read back and check that the selected scheduler is now the default.
 	 */
+	memset(tmp, 0, sizeof(tmp));
 	ret = fread(tmp, sizeof(tmp), 1, f);
 	if (ferror(f) || ret < 0) {
 		td_verror(td, errno, "fread");
 		fclose(f);
 		return 1;
 	}
-	tmp[sizeof(tmp) - 1] = '\0';
+	/*
+	 * either a list of io schedulers or "none\n" is expected.
+	 */
+	tmp[strlen(tmp) - 1] = '\0';
 
 
 	sprintf(tmp2, "[%s]", td->o.ioscheduler);
@@ -1197,20 +1261,20 @@ static int switch_ioscheduler(struct thread_data *td)
 	return 0;
 }
 
-static int keep_running(struct thread_data *td)
+static bool keep_running(struct thread_data *td)
 {
 	unsigned long long limit;
 
 	if (td->done)
-		return 0;
+		return false;
 	if (td->o.time_based)
-		return 1;
+		return true;
 	if (td->o.loops) {
 		td->o.loops--;
-		return 1;
+		return true;
 	}
 	if (exceeds_number_ios(td))
-		return 0;
+		return false;
 
 	if (td->o.io_limit)
 		limit = td->o.io_limit;
@@ -1226,20 +1290,21 @@ static int keep_running(struct thread_data *td)
 		 */
 		diff = limit - ddir_rw_sum(td->io_bytes);
 		if (diff < td_max_bs(td))
-			return 0;
+			return false;
 
 		if (fio_files_done(td))
-			return 0;
+			return false;
 
-		return 1;
+		return true;
 	}
 
-	return 0;
+	return false;
 }
 
 static int exec_string(struct thread_options *o, const char *string, const char *mode)
 {
-	int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
+	size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
+	int ret;
 	char *str;
 
 	str = malloc(newlen);
@@ -1273,7 +1338,7 @@ static uint64_t do_dry_run(struct thread_data *td)
 		if (!io_u)
 			break;
 
-		io_u->flags |= IO_U_F_FLIGHT;
+		io_u_set(io_u, IO_U_F_FLIGHT);
 		io_u->error = 0;
 		io_u->resid = 0;
 		if (ddir_rw(acct_ddir(io_u)))
@@ -1296,13 +1361,234 @@ static uint64_t do_dry_run(struct thread_data *td)
 	return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
 }
 
+static void io_workqueue_fn(struct submit_worker *sw,
+			    struct workqueue_work *work)
+{
+	struct io_u *io_u = container_of(work, struct io_u, work);
+	const enum fio_ddir ddir = io_u->ddir;
+	struct thread_data *td = sw->private;
+	int ret;
+
+	dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
+
+	io_u_set(io_u, IO_U_F_NO_FILE_PUT);
+
+	td->cur_depth++;
+
+	do {
+		ret = td_io_queue(td, io_u);
+		if (ret != FIO_Q_BUSY)
+			break;
+		ret = io_u_queued_complete(td, 1);
+		if (ret > 0)
+			td->cur_depth -= ret;
+		io_u_clear(io_u, IO_U_F_FLIGHT);
+	} while (1);
+
+	dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
+
+	io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
+
+	if (ret == FIO_Q_COMPLETED)
+		td->cur_depth--;
+	else if (ret == FIO_Q_QUEUED) {
+		unsigned int min_evts;
+
+		if (td->o.iodepth == 1)
+			min_evts = 1;
+		else
+			min_evts = 0;
+
+		ret = io_u_queued_complete(td, min_evts);
+		if (ret > 0)
+			td->cur_depth -= ret;
+	} else if (ret == FIO_Q_BUSY) {
+		ret = io_u_queued_complete(td, td->cur_depth);
+		if (ret > 0)
+			td->cur_depth -= ret;
+	}
+}
+
+static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
+{
+	struct thread_data *td = sw->private;
+
+	if (td->io_u_queued || td->cur_depth || td->io_u_in_flight)
+		return true;
+
+	return false;
+}
+
+static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
+{
+	struct thread_data *td = sw->private;
+	int ret;
+
+	ret = io_u_quiesce(td);
+	if (ret > 0)
+		td->cur_depth -= ret;
+}
+
+static int io_workqueue_alloc_fn(struct submit_worker *sw)
+{
+	struct thread_data *td;
+
+	td = calloc(1, sizeof(*td));
+	sw->private = td;
+	return 0;
+}
+
+static void io_workqueue_free_fn(struct submit_worker *sw)
+{
+	free(sw->private);
+	sw->private = NULL;
+}
+
+static int io_workqueue_init_worker_fn(struct submit_worker *sw)
+{
+	struct thread_data *parent = sw->wq->td;
+	struct thread_data *td = sw->private;
+	int fio_unused ret;
+
+	memcpy(&td->o, &parent->o, sizeof(td->o));
+	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
+	td->o.uid = td->o.gid = -1U;
+	dup_files(td, parent);
+	td->eo = parent->eo;
+	fio_options_mem_dupe(td);
+
+	if (ioengine_load(td))
+		goto err;
+
+	if (td->o.odirect)
+		td->io_ops->flags |= FIO_RAWIO;
+
+	td->pid = gettid();
+
+	INIT_FLIST_HEAD(&td->io_log_list);
+	INIT_FLIST_HEAD(&td->io_hist_list);
+	INIT_FLIST_HEAD(&td->verify_list);
+	INIT_FLIST_HEAD(&td->trim_list);
+	INIT_FLIST_HEAD(&td->next_rand_list);
+	td->io_hist_tree = RB_ROOT;
+
+	td->o.iodepth = 1;
+	if (td_io_init(td))
+		goto err_io_init;
+
+	fio_gettime(&td->epoch, NULL);
+	fio_getrusage(&td->ru_start);
+	clear_io_state(td, 1);
+
+	td_set_runstate(td, TD_RUNNING);
+	td->flags |= TD_F_CHILD;
+	td->parent = parent;
+	return 0;
+
+err_io_init:
+	close_ioengine(td);
+err:
+	return 1;
+
+}
+
+static void io_workqueue_exit_worker_fn(struct submit_worker *sw)
+{
+	struct thread_data *td = sw->private;
+
+	fio_options_free(td);
+	close_and_free_files(td);
+	if (td->io_ops)
+		close_ioengine(td);
+	td_set_runstate(td, TD_EXITED);
+}
+
+#ifdef CONFIG_SFAA
+static void sum_val(uint64_t *dst, uint64_t *src)
+{
+	if (*src) {
+		__sync_fetch_and_add(dst, *src);
+		*src = 0;
+	}
+}
+#else
+static void sum_val(uint64_t *dst, uint64_t *src)
+{
+	if (*src) {
+		*dst += *src;
+		*src = 0;
+	}
+}
+#endif
+
+static void pthread_double_unlock(pthread_mutex_t *lock1,
+				  pthread_mutex_t *lock2)
+{
+#ifndef CONFIG_SFAA
+	pthread_mutex_unlock(lock1);
+	pthread_mutex_unlock(lock2);
+#endif
+}
+
+static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
+{
+#ifndef CONFIG_SFAA
+	if (lock1 < lock2) {
+		pthread_mutex_lock(lock1);
+		pthread_mutex_lock(lock2);
+	} else {
+		pthread_mutex_lock(lock2);
+		pthread_mutex_lock(lock1);
+	}
+#endif
+}
+
+static void sum_ddir(struct thread_data *dst, struct thread_data *src,
+		     enum fio_ddir ddir)
+{
+	pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
+
+	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
+	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
+	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
+	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
+	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
+
+	pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
+}
+
+static void io_workqueue_update_acct_fn(struct submit_worker *sw)
+{
+	struct thread_data *src = sw->private;
+	struct thread_data *dst = sw->wq->td;
+
+	if (td_read(src))
+		sum_ddir(dst, src, DDIR_READ);
+	if (td_write(src))
+		sum_ddir(dst, src, DDIR_WRITE);
+	if (td_trim(src))
+		sum_ddir(dst, src, DDIR_TRIM);
+
+}
+
+struct workqueue_ops rated_wq_ops = {
+	.fn			= io_workqueue_fn,
+	.pre_sleep_flush_fn	= io_workqueue_pre_sleep_flush_fn,
+	.pre_sleep_fn		= io_workqueue_pre_sleep_fn,
+	.update_acct_fn		= io_workqueue_update_acct_fn,
+	.alloc_worker_fn	= io_workqueue_alloc_fn,
+	.free_worker_fn		= io_workqueue_free_fn,
+	.init_worker_fn		= io_workqueue_init_worker_fn,
+	.exit_worker_fn		= io_workqueue_exit_worker_fn,
+};
+
 /*
  * Entry point for the thread based jobs. The process based jobs end up
  * here as well, after a little setup.
  */
 static void *thread_main(void *data)
 {
-	unsigned long long elapsed;
+	unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
 	struct thread_data *td = data;
 	struct thread_options *o = &td->o;
 	pthread_condattr_t attr;
@@ -1493,29 +1779,34 @@ static void *thread_main(void *data)
 
 	fio_verify_init(td);
 
+	if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
+	    workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth))
+		goto err;
+
 	fio_gettime(&td->epoch, NULL);
 	fio_getrusage(&td->ru_start);
+	memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
+	memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
+
+	if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
+			o->ratemin[DDIR_TRIM]) {
+	        memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
+					sizeof(td->bw_sample_time));
+	        memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
+					sizeof(td->bw_sample_time));
+	        memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
+					sizeof(td->bw_sample_time));
+	}
+
 	clear_state = 0;
 	while (keep_running(td)) {
 		uint64_t verify_bytes;
 
 		fio_gettime(&td->start, NULL);
-		memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
-		memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
 		memcpy(&td->tv_cache, &td->start, sizeof(td->start));
 
-		if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
-				o->ratemin[DDIR_TRIM]) {
-		        memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
-						sizeof(td->bw_sample_time));
-		        memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
-						sizeof(td->bw_sample_time));
-		        memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
-						sizeof(td->bw_sample_time));
-		}
-
 		if (clear_state)
-			clear_io_state(td);
+			clear_io_state(td, 0);
 
 		prune_io_piece_log(td);
 
@@ -1536,18 +1827,12 @@ static void *thread_main(void *data)
 		check_update_rusage(td);
 
 		fio_mutex_down(stat_mutex);
-		if (td_read(td) && td->io_bytes[DDIR_READ]) {
-			elapsed = mtime_since_now(&td->start);
-			td->ts.runtime[DDIR_READ] += elapsed;
-		}
-		if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
-			elapsed = mtime_since_now(&td->start);
-			td->ts.runtime[DDIR_WRITE] += elapsed;
-		}
-		if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
-			elapsed = mtime_since_now(&td->start);
-			td->ts.runtime[DDIR_TRIM] += elapsed;
-		}
+		if (td_read(td) && td->io_bytes[DDIR_READ])
+			update_runtime(td, elapsed_us, DDIR_READ);
+		if (td_write(td) && td->io_bytes[DDIR_WRITE])
+			update_runtime(td, elapsed_us, DDIR_WRITE);
+		if (td_trim(td) && td->io_bytes[DDIR_TRIM])
+			update_runtime(td, elapsed_us, DDIR_TRIM);
 		fio_gettime(&td->start, NULL);
 		fio_mutex_up(stat_mutex);
 
@@ -1559,7 +1844,7 @@ static void *thread_main(void *data)
 		    (td->io_ops->flags & FIO_UNIDIR))
 			continue;
 
-		clear_io_state(td);
+		clear_io_state(td, 0);
 
 		fio_gettime(&td->start, NULL);
 
@@ -1571,7 +1856,7 @@ static void *thread_main(void *data)
 		check_update_rusage(td);
 
 		fio_mutex_down(stat_mutex);
-		td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
+		update_runtime(td, elapsed_us, DDIR_READ);
 		fio_gettime(&td->start, NULL);
 		fio_mutex_up(stat_mutex);
 
@@ -1586,21 +1871,16 @@ static void *thread_main(void *data)
 	td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
 
 	if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
-	    (td->o.verify != VERIFY_NONE && td_write(td))) {
-		struct all_io_list *state;
-		size_t sz;
-
-		state = get_all_io_list(td->thread_number, &sz);
-		if (state) {
-			__verify_save_state(state, "local");
-			free(state);
-		}
-	}
+	    (td->o.verify != VERIFY_NONE && td_write(td)))
+		verify_save_state(td->thread_number);
 
 	fio_unpin_memory(td);
 
 	fio_writeout_logs(td);
 
+	if (o->io_submit_mode == IO_MODE_OFFLOAD)
+		workqueue_exit(&td->io_wq);
+
 	if (td->flags & TD_F_COMPRESS_LOG)
 		tp_exit(&td->tp_data);
 
@@ -1799,29 +2079,29 @@ reaped:
 		fio_terminate_threads(TERMINATE_ALL);
 }
 
-static int __check_trigger_file(void)
+static bool __check_trigger_file(void)
 {
 	struct stat sb;
 
 	if (!trigger_file)
-		return 0;
+		return false;
 
 	if (stat(trigger_file, &sb))
-		return 0;
+		return false;
 
 	if (unlink(trigger_file) < 0)
 		log_err("fio: failed to unlink %s: %s\n", trigger_file,
 							strerror(errno));
 
-	return 1;
+	return true;
 }
 
-static int trigger_timedout(void)
+static bool trigger_timedout(void)
 {
 	if (trigger_timeout)
 		return time_since_genesis() >= trigger_timeout;
 
-	return 0;
+	return false;
 }
 
 void exec_trigger(const char *cmd)
@@ -1842,7 +2122,7 @@ void check_trigger_file(void)
 		if (nr_clients)
 			fio_clients_send_trigger(trigger_remote_cmd);
 		else {
-			verify_save_state();
+			verify_save_state(IO_LIST_ALL);
 			fio_terminate_threads(TERMINATE_ALL);
 			exec_trigger(trigger_cmd);
 		}
@@ -1858,11 +2138,12 @@ static int fio_verify_load_state(struct thread_data *td)
 
 	if (is_backend) {
 		void *data;
+		int ver;
 
 		ret = fio_server_get_verify_state(td->o.name,
-					td->thread_number - 1, &data);
+					td->thread_number - 1, &data, &ver);
 		if (!ret)
-			verify_convert_assign_state(td, data);
+			verify_convert_assign_state(td, data, ver);
 	} else
 		ret = verify_load_state(td, "local");
 
@@ -1876,6 +2157,27 @@ static void do_usleep(unsigned int usecs)
 	usleep(usecs);
 }
 
+static bool check_mount_writes(struct thread_data *td)
+{
+	struct fio_file *f;
+	unsigned int i;
+
+	if (!td_write(td) || td->o.allow_mounted_write)
+		return false;
+
+	for_each_file(td, f, i) {
+		if (f->filetype != FIO_TYPE_BD)
+			continue;
+		if (device_is_mounted(f->file_name))
+			goto mounted;
+	}
+
+	return false;
+mounted:
+	log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
+	return true;
+}
+
 /*
  * Main function for kicking off and reaping jobs, as needed.
  */
@@ -1894,13 +2196,15 @@ static void run_threads(void)
 
 	nr_thread = nr_process = 0;
 	for_each_td(td, i) {
+		if (check_mount_writes(td))
+			return;
 		if (td->o.use_thread)
 			nr_thread++;
 		else
 			nr_process++;
 	}
 
-	if (output_format == FIO_OUTPUT_NORMAL) {
+	if (output_format & FIO_OUTPUT_NORMAL) {
 		log_info("Starting ");
 		if (nr_thread)
 			log_info("%d thread%s", nr_thread,
@@ -2242,7 +2546,7 @@ int fio_backend(void)
 			for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 				struct io_log *log = agg_io_log[i];
 
-				flush_log(log);
+				flush_log(log, 0);
 				free_log(log);
 			}
 		}