X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=backend.c;h=aa94acf2499808e4543922f26ea83d21ebff06d4;hp=f82b83f75ee4d7aa09e99079d2a427c531506a2f;hb=6de659592dbebd641b262f18401c90f05db29fa1;hpb=e9d512d8e7d2b3aaeb901b4cd68793b2638e200e

diff --git a/backend.c b/backend.c
index f82b83f7..aa94acf2 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
@@ -54,6 +55,8 @@
 #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();
 }
@@ -177,7 +180,7 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
 		if (spent < td->o.ratecycle)
 			return 0;
 
-		if (td->o.rate[ddir]) {
+		if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 			/*
 			 * check bandwidth specified rate
 			 */
@@ -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 1;
 				}
 			}
 		}
@@ -362,6 +366,23 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 	return 0;
 }
 
+/*
+ * 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 int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 			       int *retptr)
 {
@@ -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) ||
@@ -520,7 +541,7 @@ sync_done:
 			*ret = ret2;
 		break;
 	default:
-		assert(ret < 0);
+		assert(*ret < 0);
 		td_verror(td, -(*ret), "td_io_queue");
 		break;
 	}
@@ -531,6 +552,12 @@ sync_done:
 	return 0;
 }
 
+static inline int 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.
@@ -623,7 +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;
@@ -664,7 +691,7 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 		 */
 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)
@@ -743,6 +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.
@@ -868,21 +927,34 @@ 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);
+		if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
+			if (td->error)
+				break;
+			ret = workqueue_enqueue(&td->io_wq, io_u);
 
-		if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 1, &comp_time))
-			break;
+			if (should_check_rate(td))
+				td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 
-		/*
-		 * See if we need to complete some commands. Note that we
-		 * can get BUSY even without IO queued, if the system is
-		 * resource starved.
-		 */
+		} else {
+			ret = td_io_queue(td, io_u);
+
+			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;
+
+			/*
+			 * 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) &&
@@ -931,7 +1003,12 @@ 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)
@@ -1145,13 +1222,17 @@ 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);
@@ -1208,7 +1289,8 @@ static int keep_running(struct thread_data *td)
 
 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);
@@ -1242,7 +1324,7 @@ 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)))
@@ -1265,13 +1347,36 @@ 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 thread_data *td, struct io_u *io_u)
+{
+	const enum fio_ddir ddir = io_u->ddir;
+	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++;
+
+	ret = td_io_queue(td, io_u);
+
+	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_QUEUED)
+		ret = io_u_queued_complete(td, 1);
+
+	td->cur_depth--;
+}
+
 /*
  * 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;
@@ -1462,29 +1567,34 @@ static void *thread_main(void *data)
 
 	fio_verify_init(td);
 
+	if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
+	    workqueue_init(td, &td->io_wq, io_workqueue_fn, 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);
 
@@ -1505,18 +1615,12 @@ 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);
 
@@ -1528,7 +1632,7 @@ 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);
 
@@ -1540,7 +1644,7 @@ 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);
 
@@ -1555,21 +1659,16 @@ 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);
 
@@ -1811,7 +1910,7 @@ 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);
 		}
@@ -1827,11 +1926,12 @@ 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");
 
@@ -1845,6 +1945,27 @@ static void do_usleep(unsigned int usecs)
 	usleep(usecs);
 }
 
+static int check_mount_writes(struct thread_data *td)
+{
+	struct fio_file *f;
+	unsigned int i;
+
+	if (!td_write(td) || td->o.allow_mounted_write)
+		return 0;
+
+	for_each_file(td, f, i) {
+		if (f->filetype != FIO_TYPE_BD)
+			continue;
+		if (device_is_mounted(f->file_name))
+			goto mounted;
+	}
+
+	return 0;
+mounted:
+	log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
+	return 1;
+}
+
 /*
  * Main function for kicking off and reaping jobs, as needed.
  */
@@ -1863,13 +1984,15 @@ 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,
@@ -2211,7 +2334,7 @@ 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);
 			}
 		}