X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=fio.c;h=478ef28aa7e2461365f2c30775e62949eb811400;hp=48d65e0db6aa1d565c7719cb7a6a177352dcbd4c;hb=127f686514ef8b1d9ab6ed702971b205454974e6;hpb=756867bd127cfa1fb27ae2d4ba973af8cb341f13

diff --git a/fio.c b/fio.c
index 48d65e0d..478ef28a 100644
--- a/fio.c
+++ b/fio.c
@@ -37,16 +37,19 @@
 #include "fio.h"
 #include "os.h"
 
-static unsigned long page_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;
 
@@ -100,13 +103,14 @@ 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;
 
 	/*
 	 * No minimum rate set, always ok
 	 */
-	if (!td->ratemin)
+	if (!td->ratemin && !td->rate_iops_min)
 		return 0;
 
 	/*
@@ -115,32 +119,55 @@ 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))
+	if (td_read(td)) {
+		iops += td->io_blocks[DDIR_READ];
 		bytes += td->this_io_bytes[DDIR_READ];
-	if (td_write(td))
+	}
+	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)
 			return 0;
 
-		if (bytes < td->rate_bytes) {
-			fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
-			return 1;
+		if (td->rate) {
+			/*
+			 * check bandwidth specified rate
+			 */
+			if (bytes < td->rate_bytes) {
+				log_err("%s: min rate %u not met\n", td->name, td->ratemin);
+				return 1;
+			} else {
+				rate = (bytes - td->rate_bytes) / spent;
+				if (rate < td->ratemin || bytes < td->rate_bytes) {
+					log_err("%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
+					return 1;
+				}
+			}
 		} else {
-			rate = (bytes - td->rate_bytes) / spent;
-			if (rate < td->ratemin || bytes < td->rate_bytes) {
-				fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
+			/*
+			 * checks iops specified rate
+			 */
+			if (iops < td->rate_iops) {
+				log_err("%s: min iops rate %u not met\n", td->name, td->rate_iops);
 				return 1;
+			} else {
+				rate = (iops - td->rate_blocks) / spent;
+				if (rate < td->rate_iops_min || iops < td->rate_blocks) {
+					log_err("%s: min iops rate %u not met, got %lu\n", td->name, td->rate_iops_min, rate);
+				}
 			}
 		}
 	}
 
 	td->rate_bytes = bytes;
+	td->rate_blocks = iops;
 	memcpy(&td->lastrate, now, sizeof(*now));
 	return 0;
 }
@@ -253,13 +280,16 @@ static void do_verify(struct thread_data *td)
 {
 	struct fio_file *f;
 	struct io_u *io_u;
-	int ret, i, min_events;
+	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) {
+		if (!(f->flags & FIO_FILE_OPEN))
+			continue;
 		if (fio_io_sync(td, f))
 			break;
 		if (file_invalidate_cache(td, f))
@@ -301,9 +331,17 @@ static void do_verify(struct thread_data *td)
 		case FIO_Q_COMPLETED:
 			if (io_u->error)
 				ret = -io_u->error;
-			else if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
+			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);
@@ -361,31 +399,6 @@ static void do_verify(struct thread_data *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.
@@ -394,7 +407,8 @@ static void do_io(struct thread_data *td)
 {
 	struct timeval s;
 	unsigned long usec;
-	int i, ret = 0;
+	unsigned int i;
+	int ret = 0;
 
 	td_set_runstate(td, TD_RUNNING);
 
@@ -424,9 +438,18 @@ static void do_io(struct thread_data *td)
 		case FIO_Q_COMPLETED:
 			if (io_u->error)
 				ret = -io_u->error;
-			else if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
+			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);
@@ -524,8 +547,12 @@ static void do_io(struct thread_data *td)
 
 		if (should_fsync(td) && td->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);
@@ -566,9 +593,6 @@ static int init_io_u(struct thread_data *td)
 	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
@@ -611,7 +635,7 @@ 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);
@@ -656,27 +680,31 @@ 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->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->last_was_sync = 0;
 
-	for_each_file(td, f, i) {
-		f->last_completed_pos = 0;
+	for_each_file(td, f, i)
+		td_io_close_file(td, f);
 
-		f->last_pos = 0;
-		if (td->io_ops->flags & FIO_SYNCIO)
-			lseek(f->fd, SEEK_SET, 0);
-
-		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;
 }
 
 /*
@@ -687,6 +715,7 @@ static void *thread_main(void *data)
 {
 	unsigned long long runtime[2];
 	struct thread_data *td = data;
+	int clear_state;
 
 	if (!td->use_thread)
 		setsid();
@@ -700,50 +729,50 @@ static void *thread_main(void *data)
 	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, "cpu_set_affinity");
-		goto err;
+		goto err_sem;
 	}
 
 	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, "ioprio_set");
-			goto err;
+			goto err_sem;
 		}
 	}
 
 	if (nice(td->nice) == -1) {
 		td_verror(td, errno, "nice");
-		goto err;
+		goto err_sem;
 	}
 
-	if (init_random_state(td))
-		goto err;
-
 	if (td->ioscheduler && switch_ioscheduler(td))
-		goto err;
+		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->create_serialize && setup_files(td))
 		goto err;
-	if (open_files(td))
-		goto err;
 
-	/*
-	 * Do this late, as some IO engines would like to have the
-	 * files setup prior to initializing structures.
-	 */
 	if (td_io_init(td))
 		goto err;
 
+	if (open_files(td))
+		goto err;
+
 	if (td->exec_prerun) {
 		if (system(td->exec_prerun) < 0)
 			goto err;
@@ -754,6 +783,7 @@ static void *thread_main(void *data)
 	getrusage(RUSAGE_SELF, &td->ts.ru_start);
 
 	runtime[0] = runtime[1] = 0;
+	clear_state = 0;
 	while (td->loops--) {
 		fio_gettime(&td->start, NULL);
 		memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
@@ -761,13 +791,14 @@ static void *thread_main(void *data)
 		if (td->ratemin)
 			memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 
-		clear_io_state(td);
+		if (clear_state && clear_io_state(td))
+			break;
+
 		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;
 
 		if (td_read(td) && td->io_bytes[DDIR_READ])
 			runtime[DDIR_READ] += utime_since_now(&td->start);
@@ -780,7 +811,9 @@ static void *thread_main(void *data)
 		if (td->verify == VERIFY_NONE)
 			continue;
 
-		clear_io_state(td);
+		if (clear_io_state(td))
+			break;
+
 		fio_gettime(&td->start, NULL);
 
 		do_verify(td);
@@ -822,6 +855,9 @@ err:
 	cleanup_io_u(td);
 	td_set_runstate(td, TD_EXITED);
 	return (void *) (unsigned long) td->error;
+err_sem:
+	fio_sem_up(startup_sem);
+	goto err;
 }
 
 /*
@@ -866,7 +902,7 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 		 * ->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->pid || td->runstate == TD_REAPED)
@@ -949,7 +985,15 @@ 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);
 	}
 
@@ -1024,23 +1068,22 @@ 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 (pthread_create(&td->thread, NULL, thread_main, td)) {
 					perror("thread_create");
 					nr_started--;
+					break;
 				}
 			} else {
-				if (fork())
-					fio_sem_down(&startup_sem);
-				else {
+				if (!fork()) {
 					int ret = fork_main(shm_id, i);
 
 					exit(ret);
 				}
 			}
+			fio_sem_down(startup_sem);
 		}
 
 		/*
@@ -1095,7 +1138,7 @@ static void run_threads(void)
 			m_rate += td->ratemin;
 			t_rate += td->rate;
 			todo--;
-			fio_sem_up(&td->mutex);
+			fio_sem_up(td->mutex);
 		}
 
 		reap_threads(&nr_running, &t_rate, &m_rate);
@@ -1138,6 +1181,7 @@ int main(int argc, char *argv[])
 		return 1;
 	}
 
+	page_size = ps;
 	page_mask = ps - 1;
 
 	if (write_bw_log) {
@@ -1145,6 +1189,8 @@ int main(int argc, char *argv[])
 		setup_log(&agg_io_log[DDIR_WRITE]);
 	}
 
+	startup_sem = fio_sem_init(0);
+
 	set_genesis_time();
 
 	disk_util_timer_arm();
@@ -1159,5 +1205,6 @@ int main(int argc, char *argv[])
 		}
 	}
 
+	fio_sem_remove(startup_sem);
 	return exit_value;
 }