X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=fio.c;h=ebe06d66f1ed1131e472ff2d0515acf6ed7b443c;hp=5a87ae4b000fe377ad9970fa7971f725c287c020;hb=581e71417760e0aa86eac0acd704253ff0eeea4f;hpb=e3cedca76d9fc104eb4f6f869606fb5bf4c0d59c

diff --git a/fio.c b/fio.c
index 5a87ae4b..ebe06d66 100644
--- a/fio.c
+++ b/fio.c
@@ -37,6 +37,8 @@
 #include "fio.h"
 #include "hash.h"
 #include "smalloc.h"
+#include "verify.h"
+#include "diskutil.h"
 
 unsigned long page_mask;
 unsigned long page_size;
@@ -52,9 +54,11 @@ int temp_stall_ts;
 unsigned long done_secs = 0;
 
 static struct fio_mutex *startup_mutex;
+static struct fio_mutex *writeout_mutex;
 static volatile int fio_abort;
 static int exit_value;
 static struct itimerval itimer;
+static pthread_t gtod_thread;
 
 struct io_log *agg_io_log[2];
 
@@ -168,26 +172,19 @@ static void set_sig_handlers(void)
 	sigaction(SIGQUIT, &act, NULL);
 }
 
-static inline int should_check_rate(struct thread_data *td)
-{
-	/*
-	 * No minimum rate set, always ok
-	 */
-	if (!td->o.ratemin && !td->o.rate_iops_min)
-		return 0;
-
-	return 1;
-}
-
 /*
  * Check if we are above the minimum rate given.
  */
-static int check_min_rate(struct thread_data *td, struct timeval *now)
+static int __check_min_rate(struct thread_data *td, struct timeval *now,
+			    enum td_ddir ddir)
 {
 	unsigned long long bytes = 0;
 	unsigned long iops = 0;
 	unsigned long spent;
 	unsigned long rate;
+	unsigned int ratemin = 0;
+	unsigned int rate_iops = 0;
+	unsigned int rate_iops_min = 0;
 
 	/*
 	 * allow a 2 second settle period in the beginning
@@ -195,38 +192,35 @@ 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];
-	}
+	iops += td->io_blocks[ddir];
+	bytes += td->this_io_bytes[ddir];
+	ratemin += td->o.ratemin[ddir];
+	rate_iops += td->o.rate_iops[ddir];
+	rate_iops_min += td->o.rate_iops_min[ddir];
 
 	/*
 	 * if rate blocks is set, sample is running
 	 */
-	if (td->rate_bytes || td->rate_blocks) {
-		spent = mtime_since(&td->lastrate, now);
+	if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
+		spent = mtime_since(&td->lastrate[ddir], now);
 		if (spent < td->o.ratecycle)
 			return 0;
 
-		if (td->o.rate) {
+		if (td->o.rate[ddir]) {
 			/*
 			 * check bandwidth specified rate
 			 */
-			if (bytes < td->rate_bytes) {
+			if (bytes < td->rate_bytes[ddir]) {
 				log_err("%s: min rate %u not met\n", td->o.name,
-								td->o.ratemin);
+								ratemin);
 				return 1;
 			} else {
-				rate = (bytes - td->rate_bytes) / spent;
-				if (rate < td->o.ratemin ||
-				    bytes < td->rate_bytes) {
+				rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
+				if (rate < ratemin ||
+				    bytes < td->rate_bytes[ddir]) {
 					log_err("%s: min rate %u not met, got"
 						" %luKiB/sec\n", td->o.name,
-							td->o.ratemin, rate);
+							ratemin, rate);
 					return 1;
 				}
 			}
@@ -234,29 +228,41 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
 			/*
 			 * checks iops specified rate
 			 */
-			if (iops < td->o.rate_iops) {
+			if (iops < rate_iops) {
 				log_err("%s: min iops rate %u not met\n",
-						td->o.name, td->o.rate_iops);
+						td->o.name, rate_iops);
 				return 1;
 			} else {
-				rate = (iops - td->rate_blocks) / spent;
-				if (rate < td->o.rate_iops_min ||
-				    iops < td->rate_blocks) {
+				rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
+				if (rate < rate_iops_min ||
+				    iops < td->rate_blocks[ddir]) {
 					log_err("%s: min iops rate %u not met,"
 						" got %lu\n", td->o.name,
-							td->o.rate_iops_min,
-							rate);
+							rate_iops_min, rate);
 				}
 			}
 		}
 	}
 
-	td->rate_bytes = bytes;
-	td->rate_blocks = iops;
-	memcpy(&td->lastrate, now, sizeof(*now));
+	td->rate_bytes[ddir] = bytes;
+	td->rate_blocks[ddir] = iops;
+	memcpy(&td->lastrate[ddir], now, sizeof(*now));
 	return 0;
 }
 
+static int check_min_rate(struct thread_data *td, struct timeval *now,
+			  unsigned long *bytes_done)
+{
+	int ret = 0;
+
+	if (bytes_done[0])
+		ret |= __check_min_rate(td, now, 0);
+	if (bytes_done[1])
+		ret |= __check_min_rate(td, now, 1);
+
+	return ret;
+}
+
 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 {
 	if (!td->o.timeout)
@@ -280,7 +286,7 @@ static void cleanup_pending_aio(struct thread_data *td)
 	/*
 	 * get immediately available events, if any
 	 */
-	r = io_u_queued_complete(td, 0);
+	r = io_u_queued_complete(td, 0, NULL);
 	if (r < 0)
 		return;
 
@@ -308,7 +314,7 @@ static void cleanup_pending_aio(struct thread_data *td)
 	}
 
 	if (td->cur_depth)
-		r = io_u_queued_complete(td, td->cur_depth);
+		r = io_u_queued_complete(td, td->cur_depth, NULL);
 }
 
 /*
@@ -338,7 +344,7 @@ requeue:
 		put_io_u(td, io_u);
 		return 1;
 	} else if (ret == FIO_Q_QUEUED) {
-		if (io_u_queued_complete(td, 1) < 0)
+		if (io_u_queued_complete(td, 1, NULL) < 0)
 			return 1;
 	} else if (ret == FIO_Q_COMPLETED) {
 		if (io_u->error) {
@@ -346,7 +352,7 @@ requeue:
 			return 1;
 		}
 
-		if (io_u_sync_complete(td, io_u) < 0)
+		if (io_u_sync_complete(td, io_u, NULL) < 0)
 			return 1;
 	} else if (ret == FIO_Q_BUSY) {
 		if (td_io_commit(td))
@@ -379,7 +385,7 @@ static void do_verify(struct thread_data *td)
 	 * read from disk.
 	 */
 	for_each_file(td, f, i) {
-		if (!(f->flags & FIO_FILE_OPEN))
+		if (!fio_file_open(f))
 			continue;
 		if (fio_io_sync(td, f))
 			break;
@@ -450,7 +456,7 @@ static void do_verify(struct thread_data *td)
 				requeue_io_u(td, &io_u);
 			} else {
 sync_done:
-				ret = io_u_sync_complete(td, io_u);
+				ret = io_u_sync_complete(td, io_u, NULL);
 				if (ret < 0)
 					break;
 			}
@@ -488,7 +494,7 @@ sync_done:
 				 * and do the verification on them through
 				 * the callback handler
 				 */
-				if (io_u_queued_complete(td, min_events) < 0) {
+				if (io_u_queued_complete(td, min_events, NULL) < 0) {
 					ret = -1;
 					break;
 				}
@@ -502,7 +508,7 @@ sync_done:
 		min_events = td->cur_depth;
 
 		if (min_events)
-			ret = io_u_queued_complete(td, min_events);
+			ret = io_u_queued_complete(td, min_events, NULL);
 	} else
 		cleanup_pending_aio(td);
 
@@ -515,7 +521,6 @@ sync_done:
  */
 static void do_io(struct thread_data *td)
 {
-	unsigned long usec;
 	unsigned int i;
 	int ret = 0;
 
@@ -526,7 +531,7 @@ static void do_io(struct thread_data *td)
 
 	while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 		struct timeval comp_time;
-		long bytes_done = 0;
+		unsigned long bytes_done[2] = { 0, 0 };
 		int min_evts = 0;
 		struct io_u *io_u;
 		int ret2, full;
@@ -588,12 +593,13 @@ static void do_io(struct thread_data *td)
 				requeue_io_u(td, &io_u);
 			} else {
 sync_done:
-				if (should_check_rate(td))
+				if (__should_check_rate(td, 0) ||
+				    __should_check_rate(td, 1))
 					fio_gettime(&comp_time, NULL);
 
-				bytes_done = io_u_sync_complete(td, io_u);
-				if (bytes_done < 0)
-					ret = bytes_done;
+				ret = io_u_sync_complete(td, io_u, bytes_done);
+				if (ret < 0)
+					break;
 			}
 			break;
 		case FIO_Q_QUEUED:
@@ -629,34 +635,25 @@ sync_done:
 			if (full && !min_evts)
 				min_evts = 1;
 
-			if (should_check_rate(td))
+			if (__should_check_rate(td, 0) ||
+			    __should_check_rate(td, 1))
 				fio_gettime(&comp_time, NULL);
 
 			do {
-				ret = io_u_queued_complete(td, min_evts);
-				if (ret <= 0)
+				ret = io_u_queued_complete(td, min_evts, bytes_done);
+				if (ret < 0)
 					break;
 
-				bytes_done += ret;
 			} while (full && (td->cur_depth > td->o.iodepth_low));
 		}
 
 		if (ret < 0)
 			break;
-		if (!bytes_done)
+		if (!(bytes_done[0] + bytes_done[1]))
 			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
-		 */
-		if (!in_ramp_time(td) && should_check_rate(td)) {
-			usec = utime_since(&td->tv_cache, &comp_time);
-
-			rate_throttle(td, usec, bytes_done);
-
-			if (check_min_rate(td, &comp_time)) {
+		if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
+			if (check_min_rate(td, &comp_time, bytes_done)) {
 				if (exitall_on_terminate)
 					terminate_threads(td->groupid);
 				td_verror(td, EIO, "check_min_rate");
@@ -672,7 +669,7 @@ sync_done:
 				int left;
 
 				if (td->o.thinktime_spin)
-					__usec_sleep(td->o.thinktime_spin);
+					usec_spin(td->o.thinktime_spin);
 
 				left = td->o.thinktime - td->o.thinktime_spin;
 				if (left)
@@ -690,13 +687,13 @@ sync_done:
 
 		i = td->cur_depth;
 		if (i)
-			ret = io_u_queued_complete(td, i);
+			ret = io_u_queued_complete(td, i, NULL);
 
 		if (should_fsync(td) && td->o.end_fsync) {
 			td_set_runstate(td, TD_FSYNCING);
 
 			for_each_file(td, f, i) {
-				if (!(f->flags & FIO_FILE_OPEN))
+				if (!fio_file_open(f))
 					continue;
 				fio_io_sync(td, f);
 			}
@@ -730,7 +727,7 @@ static int init_io_u(struct thread_data *td)
 {
 	struct io_u *io_u;
 	unsigned int max_bs;
-	int i, max_units;
+	int cl_align, i, max_units;
 	char *p;
 
 	max_units = td->o.iodepth;
@@ -758,10 +755,20 @@ static int init_io_u(struct thread_data *td)
 	else
 		p = td->orig_buffer;
 
+	cl_align = os_cache_line_size();
+
 	for (i = 0; i < max_units; i++) {
+		void *ptr;
+
 		if (td->terminate)
 			return 1;
-		io_u = malloc(sizeof(*io_u));
+
+		if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
+			log_err("fio: posix_memalign=%s\n", strerror(errno));
+			break;
+		}
+
+		io_u = ptr;
 		memset(io_u, 0, sizeof(*io_u));
 		INIT_FLIST_HEAD(&io_u->list);
 
@@ -862,9 +869,8 @@ static void reset_io_counters(struct thread_data *td)
 	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;
+	td->rate_bytes[0] = td->rate_bytes[1] = 0;
+	td->rate_blocks[0] = td->rate_blocks[1] = 0;
 
 	td->last_was_sync = 0;
 
@@ -873,6 +879,11 @@ static void reset_io_counters(struct thread_data *td)
 	 */
 	if (td->o.time_based || td->o.loops)
 		td->nr_done_files = 0;
+
+	/*
+	 * Set the same seed to get repeatable runs
+	 */
+	td_fill_rand_seeds(td);
 }
 
 void reset_all_stats(struct thread_data *td)
@@ -894,24 +905,31 @@ void reset_all_stats(struct thread_data *td)
 	memcpy(&td->start, &tv, sizeof(tv));
 }
 
-static int clear_io_state(struct thread_data *td)
+static void clear_io_state(struct thread_data *td)
 {
 	struct fio_file *f;
 	unsigned int i;
-	int ret;
 
 	reset_io_counters(td);
 
 	close_files(td);
+	for_each_file(td, f, i)
+		fio_file_clear_done(f);
+}
 
-	ret = 0;
-	for_each_file(td, f, i) {
-		f->flags &= ~FIO_FILE_DONE;
-		ret = td_io_open_file(td, f);
-		if (ret)
-			break;
-	}
+static int exec_string(const char *string)
+{
+	int ret, newlen = strlen(string) + 1 + 8;
+	char *str;
+
+	str = malloc(newlen);
+	sprintf(str, "sh -c %s", string);
 
+	ret = system(str);
+	if (ret == -1)
+		log_err("fio: exec of cmd <%s> failed\n", str);
+
+	free(str);
 	return ret;
 }
 
@@ -940,8 +958,11 @@ static void *thread_main(void *data)
 	td->io_hist_tree = RB_ROOT;
 
 	td_set_runstate(td, TD_INITIALIZED);
+	dprint(FD_MUTEX, "up startup_mutex\n");
 	fio_mutex_up(startup_mutex);
+	dprint(FD_MUTEX, "wait on td->mutex\n");
 	fio_mutex_down(td->mutex);
+	dprint(FD_MUTEX, "done waiting on td->mutex\n");
 
 	/*
 	 * the ->mutex mutex is now no longer used, close it to avoid
@@ -964,6 +985,18 @@ static void *thread_main(void *data)
 		goto err;
 	}
 
+	/*
+	 * If we have a gettimeofday() thread, make sure we exclude that
+	 * thread from this job
+	 */
+	if (td->o.gtod_cpu) {
+		fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
+		if (fio_setaffinity(td) == -1) {
+			td_verror(td, errno, "cpu_set_affinity");
+			goto err;
+		}
+	}
+
 	if (td->ioprio_set) {
 		if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 			td_verror(td, errno, "ioprio_set");
@@ -985,14 +1018,16 @@ static void *thread_main(void *data)
 	if (td_io_init(td))
 		goto err;
 
-	if (open_files(td))
-		goto err;
-
 	if (init_random_map(td))
 		goto err;
 
 	if (td->o.exec_prerun) {
-		if (system(td->o.exec_prerun) < 0)
+		if (exec_string(td->o.exec_prerun))
+			goto err;
+	}
+
+	if (td->o.pre_read) {
+		if (pre_read_files(td) < 0)
 			goto err;
 	}
 
@@ -1010,8 +1045,8 @@ static void *thread_main(void *data)
 			memcpy(&td->lastrate, &td->ts.stat_sample_time,
 							sizeof(td->lastrate));
 
-		if (clear_state && clear_io_state(td))
-			break;
+		if (clear_state)
+			clear_io_state(td);
 
 		prune_io_piece_log(td);
 
@@ -1020,21 +1055,11 @@ static void *thread_main(void *data)
 		clear_state = 1;
 
 		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);
-
+			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);
-
+			elapsed = utime_since_now(&td->start);
 			runtime[DDIR_WRITE] += elapsed;
 		}
 
@@ -1046,8 +1071,7 @@ static void *thread_main(void *data)
 		    (td->io_ops->flags & FIO_UNIDIR))
 			continue;
 
-		if (clear_io_state(td))
-			break;
+		clear_io_state(td);
 
 		fio_gettime(&td->start, NULL);
 
@@ -1066,28 +1090,31 @@ static void *thread_main(void *data)
 	td->ts.io_bytes[0] = td->io_bytes[0];
 	td->ts.io_bytes[1] = td->io_bytes[1];
 
+	fio_mutex_down(writeout_mutex);
 	if (td->ts.bw_log) {
-		if (td->o.bw_log_file)
-			finish_log_named(td, td->ts.bw_log, td->o.bw_log_file, "bw");
-		else
+		if (td->o.bw_log_file) {
+			finish_log_named(td, td->ts.bw_log,
+						td->o.bw_log_file, "bw");
+		} else
 			finish_log(td, td->ts.bw_log, "bw");
 	}
 	if (td->ts.slat_log) {
-		if (td->o.lat_log_file)
-			finish_log_named(td, td->ts.slat_log, td->o.lat_log_file, "clat");
-		else
+		if (td->o.lat_log_file) {
+			finish_log_named(td, td->ts.slat_log,
+						td->o.lat_log_file, "slat");
+		} else
 			finish_log(td, td->ts.slat_log, "slat");
 	}
 	if (td->ts.clat_log) {
-		if (td->o.lat_log_file)
-			finish_log_named(td, td->ts.clat_log, td->o.lat_log_file, "clat");
-		else
+		if (td->o.lat_log_file) {
+			finish_log_named(td, td->ts.clat_log,
+						td->o.lat_log_file, "clat");
+		} else
 			finish_log(td, td->ts.clat_log, "clat");
 	}
-	if (td->o.exec_postrun) {
-		if (system(td->o.exec_postrun) < 0)
-			log_err("fio: postrun %s failed\n", td->o.exec_postrun);
-	}
+	fio_mutex_up(writeout_mutex);
+	if (td->o.exec_postrun)
+		exec_string(td->o.exec_postrun);
 
 	if (exitall_on_terminate)
 		terminate_threads(td->groupid);
@@ -1100,6 +1127,12 @@ err:
 	close_ioengine(td);
 	cleanup_io_u(td);
 
+	if (td->o.cpumask_set) {
+		int ret = fio_cpuset_exit(&td->o.cpumask);
+
+		td_verror(td, ret, "fio_cpuset_exit");
+	}
+
 	/*
 	 * do this very late, it will log file closing as well
 	 */
@@ -1214,8 +1247,8 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 		continue;
 reaped:
 		(*nr_running)--;
-		(*m_rate) -= td->o.ratemin;
-		(*t_rate) -= td->o.rate;
+		(*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
+		(*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
 		if (!td->pid)
 			pending--;
 
@@ -1229,6 +1262,46 @@ reaped:
 		terminate_threads(TERMINATE_ALL);
 }
 
+static void *gtod_thread_main(void *data)
+{
+	fio_mutex_up(startup_mutex);
+
+	/*
+	 * As long as we have jobs around, update the clock. It would be nice
+	 * to have some way of NOT hammering that CPU with gettimeofday(),
+	 * but I'm not sure what to use outside of a simple CPU nop to relax
+	 * it - we don't want to lose precision.
+	 */
+	while (threads) {
+		fio_gtod_update();
+		nop;
+	}
+
+	return NULL;
+}
+
+static int fio_start_gtod_thread(void)
+{
+	int ret;
+
+	ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
+	if (ret) {
+		log_err("Can't create gtod thread: %s\n", strerror(ret));
+		return 1;
+	}
+
+	ret = pthread_detach(gtod_thread);
+	if (ret) {
+		log_err("Can't detatch gtod thread: %s\n", strerror(ret));
+		return 1;
+	}
+
+	dprint(FD_MUTEX, "wait on startup_mutex\n");
+	fio_mutex_down(startup_mutex);
+	dprint(FD_MUTEX, "done waiting on startup_mutex\n");
+	return 0;
+}
+
 /*
  * Main function for kicking off and reaping jobs, as needed.
  */
@@ -1241,6 +1314,9 @@ static void run_threads(void)
 	if (fio_pin_memory())
 		return;
 
+	if (fio_gtod_offload && fio_start_gtod_thread())
+		return;
+
 	if (!terse_output) {
 		printf("Starting ");
 		if (nr_thread)
@@ -1344,15 +1420,21 @@ static void run_threads(void)
 			nr_started++;
 
 			if (td->o.use_thread) {
+				int ret;
+
 				dprint(FD_PROCESS, "will pthread_create\n");
-				if (pthread_create(&td->thread, NULL,
-						   thread_main, td)) {
-					perror("pthread_create");
+				ret = pthread_create(&td->thread, NULL,
+							thread_main, td);
+				if (ret) {
+					log_err("pthread_create: %s\n",
+							strerror(ret));
 					nr_started--;
 					break;
 				}
-				if (pthread_detach(td->thread) < 0)
-					perror("pthread_detach");
+				ret = pthread_detach(td->thread);
+				if (ret)
+					log_err("pthread_detach: %s",
+							strerror(ret));
 			} else {
 				pid_t pid;
 				dprint(FD_PROCESS, "will fork\n");
@@ -1364,7 +1446,9 @@ static void run_threads(void)
 				} else if (i == fio_debug_jobno)
 					*fio_debug_jobp = pid;
 			}
+			dprint(FD_MUTEX, "wait on startup_mutex\n");
 			fio_mutex_down(startup_mutex);
+			dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 		}
 
 		/*
@@ -1419,8 +1503,8 @@ static void run_threads(void)
 				td_set_runstate(td, TD_RUNNING);
 			nr_running++;
 			nr_started--;
-			m_rate += td->o.ratemin;
-			t_rate += td->o.rate;
+			m_rate += td->o.ratemin[0] + td->o.ratemin[1];
+			t_rate += td->o.rate[0] + td->o.rate[1];
 			todo--;
 			fio_mutex_up(td->mutex);
 		}
@@ -1474,6 +1558,7 @@ int main(int argc, char *argv[])
 	}
 
 	startup_mutex = fio_mutex_init(0);
+	writeout_mutex = fio_mutex_init(1);
 
 	set_genesis_time();
 
@@ -1491,5 +1576,6 @@ int main(int argc, char *argv[])
 	}
 
 	fio_mutex_remove(startup_mutex);
+	fio_mutex_remove(writeout_mutex);
 	return exit_value;
 }