backend.c

/*
 * fio - the flexible io tester
 *
 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
 * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
 *
 * The license below covers all files distributed with fio unless otherwise
 * noted in the file itself.
 *
 *  This program is free software; you can redistribute it and/or modify
 *  it under the terms of the GNU General Public License version 2 as
 *  published by the Free Software Foundation.
 *
 *  This program is distributed in the hope that it will be useful,
 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *  GNU General Public License for more details.
 *
 *  You should have received a copy of the GNU General Public License
 *  along with this program; if not, write to the Free Software
 *  Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
 *
 */
#include <unistd.h>
#include <string.h>
#include <signal.h>
#include <assert.h>
#include <inttypes.h>
#include <sys/stat.h>
#include <sys/wait.h>
#include <math.h>
#include <pthread.h>

#include "fio.h"
#include "smalloc.h"
#include "verify.h"
#include "diskutil.h"
#include "cgroup.h"
#include "profile.h"
#include "lib/rand.h"
#include "lib/memalign.h"
#include "server.h"
#include "lib/getrusage.h"
#include "idletime.h"
#include "err.h"
#include "workqueue.h"
#include "lib/mountcheck.h"
#include "rate-submit.h"
#include "helper_thread.h"
#include "pshared.h"
#include "zone-dist.h"
#include "fio_time.h"

static struct fio_sem *startup_sem;
static struct flist_head *cgroup_list;
static struct cgroup_mnt *cgroup_mnt;
static int exit_value;
static volatile bool fio_abort;
static unsigned int nr_process = 0;
static unsigned int nr_thread = 0;

struct io_log *agg_io_log[DDIR_RWDIR_CNT];

int groupid = 0;
unsigned int thread_number = 0;
unsigned int nr_segments = 0;
unsigned int cur_segment = 0;
unsigned int stat_number = 0;
int temp_stall_ts;
unsigned long done_secs = 0;
#ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
#else
pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
#endif

#define JOB_START_TIMEOUT	(5 * 1000)

static void sig_int(int sig)
{
	if (nr_segments) {
		if (is_backend)
			fio_server_got_signal(sig);
		else {
			log_info("\nfio: terminating on signal %d\n", sig);
			log_info_flush();
			exit_value = 128;
		}

		fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
	}
}

#ifdef WIN32
static void sig_break(int sig)
{
	sig_int(sig);

	/**
	 * Windows terminates all job processes on SIGBREAK after the handler
	 * returns, so give them time to wrap-up and give stats
	 */
	for_each_td(td) {
		while (td->runstate < TD_EXITED)
			sleep(1);
	} end_for_each();
}
#endif

void sig_show_status(int sig)
{
	show_running_run_stats();
}

static void set_sig_handlers(void)
{
	struct sigaction act;

	memset(&act, 0, sizeof(act));
	act.sa_handler = sig_int;
	act.sa_flags = SA_RESTART;
	sigaction(SIGINT, &act, NULL);

	memset(&act, 0, sizeof(act));
	act.sa_handler = sig_int;
	act.sa_flags = SA_RESTART;
	sigaction(SIGTERM, &act, NULL);

/* Windows uses SIGBREAK as a quit signal from other applications */
#ifdef WIN32
	memset(&act, 0, sizeof(act));
	act.sa_handler = sig_break;
	act.sa_flags = SA_RESTART;
	sigaction(SIGBREAK, &act, NULL);
#endif

	memset(&act, 0, sizeof(act));
	act.sa_handler = sig_show_status;
	act.sa_flags = SA_RESTART;
	sigaction(SIGUSR1, &act, NULL);

	if (is_backend) {
		memset(&act, 0, sizeof(act));
		act.sa_handler = sig_int;
		act.sa_flags = SA_RESTART;
		sigaction(SIGPIPE, &act, NULL);
	}
}

/*
 * Check if we are above the minimum rate given.
 */
static bool __check_min_rate(struct thread_data *td, struct timespec *now,
			     enum fio_ddir ddir)
{
	unsigned long long current_rate_check_bytes = td->this_io_bytes[ddir];
	unsigned long current_rate_check_blocks = td->this_io_blocks[ddir];
	unsigned long long option_rate_bytes_min = td->o.ratemin[ddir];
	unsigned int option_rate_iops_min = td->o.rate_iops_min[ddir];

	assert(ddir_rw(ddir));

	if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
		return false;

	/*
	 * allow a 2 second settle period in the beginning
	 */
	if (mtime_since(&td->start, now) < 2000)
		return false;

	/*
	 * if last_rate_check_blocks or last_rate_check_bytes is set,
	 * we can compute a rate per ratecycle
	 */
	if (td->last_rate_check_bytes[ddir] || td->last_rate_check_blocks[ddir]) {
		unsigned long spent = mtime_since(&td->last_rate_check_time[ddir], now);
		if (spent < td->o.ratecycle || spent==0)
			return false;

		if (td->o.ratemin[ddir]) {
			/*
			 * check bandwidth specified rate
			 */
			unsigned long long current_rate_bytes =
				((current_rate_check_bytes - td->last_rate_check_bytes[ddir]) * 1000) / spent;
			if (current_rate_bytes < option_rate_bytes_min) {
				log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
					td->o.name, option_rate_bytes_min, current_rate_bytes);
				return true;
			}
		} else {
			/*
			 * checks iops specified rate
			 */
			unsigned long long current_rate_iops =
				((current_rate_check_blocks - td->last_rate_check_blocks[ddir]) * 1000) / spent;

			if (current_rate_iops < option_rate_iops_min) {
				log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
					td->o.name, option_rate_iops_min, current_rate_iops);
				return true;
			}
		}
	}

	td->last_rate_check_bytes[ddir] = current_rate_check_bytes;
	td->last_rate_check_blocks[ddir] = current_rate_check_blocks;
	memcpy(&td->last_rate_check_time[ddir], now, sizeof(*now));
	return false;
}

static bool check_min_rate(struct thread_data *td, struct timespec *now)
{
	bool ret = false;

	for_each_rw_ddir(ddir) {
		if (td->bytes_done[ddir])
			ret |= __check_min_rate(td, now, ddir);
	}

	return ret;
}

/*
 * 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)
{
	int r;

	/*
	 * get immediately available events, if any
	 */
	r = io_u_queued_complete(td, 0);

	/*
	 * now cancel remaining active events
	 */
	if (td->io_ops->cancel) {
		struct io_u *io_u;
		int i;

		io_u_qiter(&td->io_u_all, io_u, i) {
			if (io_u->flags & IO_U_F_FLIGHT) {
				r = td->io_ops->cancel(td, io_u);
				if (!r)
					put_io_u(td, io_u);
			}
		}
	}

	if (td->cur_depth)
		r = io_u_queued_complete(td, td->cur_depth);
}

/*
 * Helper to handle the final sync of a file. Works just like the normal
 * io path, just does everything sync.
 */
static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
{
	struct io_u *io_u = __get_io_u(td);
	enum fio_q_status ret;

	if (!io_u)
		return true;

	io_u->ddir = DDIR_SYNC;
	io_u->file = f;
	io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);

	if (td_io_prep(td, io_u)) {
		put_io_u(td, io_u);
		return true;
	}

requeue:
	ret = td_io_queue(td, io_u);
	switch (ret) {
	case FIO_Q_QUEUED:
		td_io_commit(td);
		if (io_u_queued_complete(td, 1) < 0)
			return true;
		break;
	case FIO_Q_COMPLETED:
		if (io_u->error) {
			td_verror(td, io_u->error, "td_io_queue");
			return true;
		}

		if (io_u_sync_complete(td, io_u) < 0)
			return true;
		break;
	case FIO_Q_BUSY:
		td_io_commit(td);
		goto requeue;
	}

	return false;
}

static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
{
	int ret, ret2;

	if (fio_file_open(f))
		return fio_io_sync(td, f);

	if (td_io_open_file(td, f))
		return 1;

	ret = fio_io_sync(td, f);
	ret2 = 0;
	if (fio_file_open(f))
		ret2 = td_io_close_file(td, f);
	return (ret || ret2);
}

static inline void __update_ts_cache(struct thread_data *td)
{
	fio_gettime(&td->ts_cache, NULL);
}

static inline void update_ts_cache(struct thread_data *td)
{
	if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
		__update_ts_cache(td);
}

static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
{
	if (in_ramp_time(td))
		return false;
	if (!td->o.timeout)
		return false;
	if (utime_since(&td->epoch, t) >= td->o.timeout)
		return true;

	return false;
}

/*
 * 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;

	if (ret < 0 || td->error) {
		int err = td->error;
		enum error_type_bit eb;

		if (ret < 0)
			err = -ret;

		eb = td_error_type(ddir, err);
		if (!(td->o.continue_on_error & (1 << eb)))
			return true;

		if (td_non_fatal_error(td, eb, err)) {
		        /*
		         * Continue with the I/Os in case of
			 * a non fatal error.
			 */
			update_error_count(td, err);
			td_clear_error(td);
			*retptr = 0;
			return false;
		} else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
			/*
			 * We expect to hit this error if
			 * fill_device option is set.
			 */
			td_clear_error(td);
			fio_mark_td_terminate(td);
			return true;
		} else {
			/*
			 * Stop the I/O in case of a fatal
			 * error.
			 */
			update_error_count(td, err);
			return true;
		}
	}

	return false;
}

static void check_update_rusage(struct thread_data *td)
{
	if (td->update_rusage) {
		td->update_rusage = 0;
		update_rusage_stat(td);
		fio_sem_up(td->rusage_sem);
	}
}

static int wait_for_completions(struct thread_data *td, struct timespec *time)
{
	const int full = queue_full(td);
	int min_evts = 0;
	int ret;

	if (td->flags & TD_F_REGROW_LOGS)
		return io_u_quiesce(td);

	/*
	 * if the queue is full, we MUST reap at least 1 event
	 */
	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))
		fio_gettime(time, NULL);

	do {
		ret = io_u_queued_complete(td, min_evts);
		if (ret < 0)
			break;
	} while (full && (td->cur_depth > td->o.iodepth_low));

	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 timespec *comp_time)
{
	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) {
			long long 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(io_u);

			/*
			 * zero read, fail
			 */
			if (!bytes) {
				if (!from_verify)
					unlog_io_piece(td, io_u);
				td_verror(td, EIO, "full resid");
				clear_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]++;

			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))
				fio_gettime(comp_time, NULL);

			*ret = io_u_sync_complete(td, io_u);
			if (*ret < 0)
				break;
		}

		if (td->flags & TD_F_REGROW_LOGS)
			regrow_logs(td);

		/*
		 * when doing I/O (not when verifying),
		 * check for any errors that are to be ignored
		 */
		if (!from_verify)
			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);
		td_io_commit(td);
		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;
}
/*
 * Unlinks files from thread data fio_file structure
 */
static int unlink_all_files(struct thread_data *td)
{
	struct fio_file *f;
	unsigned int i;
	int ret = 0;

	for_each_file(td, f, i) {
		if (f->filetype != FIO_TYPE_FILE)
			continue;
		ret = td_io_unlink_file(td, f);
		if (ret)
			break;
	}

	if (ret)
		td_verror(td, ret, "unlink_all_files");

	return ret;
}

/*
 * Check if io_u will overlap an in-flight IO in the queue
 */
bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
{
	bool overlap;
	struct io_u *check_io_u;
	unsigned long long x1, x2, y1, y2;
	int i;

	x1 = io_u->offset;
	x2 = io_u->offset + io_u->buflen;
	overlap = false;
	io_u_qiter(q, check_io_u, i) {
		if (check_io_u->flags & IO_U_F_FLIGHT) {
			y1 = check_io_u->offset;
			y2 = check_io_u->offset + check_io_u->buflen;

			if (x1 < y2 && y1 < x2) {
				overlap = true;
				dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
						x1, io_u->buflen,
						y1, check_io_u->buflen);
				break;
			}
		}
	}

	return overlap;
}

static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
{
	/*
	 * Check for overlap if the user asked us to, and we have
	 * at least one IO in flight besides this one.
	 */
	if (td->o.serialize_overlap && td->cur_depth > 1 &&
	    in_flight_overlap(&td->io_u_all, io_u))
		return FIO_Q_BUSY;

	return td_io_queue(td, io_u);
}

/*
 * 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, uint64_t verify_bytes)
{
	struct fio_file *f;
	struct io_u *io_u;
	int ret, min_events;
	unsigned int i;

	dprint(FD_VERIFY, "starting loop\n");

	/*
	 * sync io first and invalidate cache, to make sure we really
	 * read from disk.
	 */
	for_each_file(td, f, i) {
		if (!fio_file_open(f))
			continue;
		if (fio_io_sync(td, f))
			break;
		if (file_invalidate_cache(td, f))
			break;
	}

	check_update_rusage(td);

	if (td->error)
		return;

	td_set_runstate(td, TD_VERIFYING);

	io_u = NULL;
	while (!td->terminate) {
		enum fio_ddir ddir;
		int full;

		update_ts_cache(td);
		check_update_rusage(td);

		if (runtime_exceeded(td, &td->ts_cache)) {
			__update_ts_cache(td);
			if (runtime_exceeded(td, &td->ts_cache)) {
				fio_mark_td_terminate(td);
				break;
			}
		}

		if (flow_threshold_exceeded(td))
			continue;

		if (!td->o.experimental_verify) {
			io_u = __get_io_u(td);
			if (!io_u)
				break;

			if (get_next_verify(td, io_u)) {
				put_io_u(td, io_u);
				break;
			}

			if (td_io_prep(td, io_u)) {
				put_io_u(td, io_u);
				break;
			}
		} else {
			if (td->bytes_verified + td->o.rw_min_bs > verify_bytes)
				break;

			while ((io_u = get_io_u(td)) != NULL) {
				if (IS_ERR_OR_NULL(io_u)) {
					io_u = NULL;
					ret = FIO_Q_BUSY;
					goto reap;
				}

				/*
				 * We are only interested in the places where
				 * we wrote or trimmed IOs. Turn those into
				 * reads for verification purposes.
				 */
				if (io_u->ddir == DDIR_READ) {
					/*
					 * Pretend we issued it for rwmix
					 * accounting
					 */
					td->io_issues[DDIR_READ]++;
					put_io_u(td, io_u);
					continue;
				} else if (io_u->ddir == DDIR_TRIM) {
					io_u->ddir = DDIR_READ;
					io_u_set(td, io_u, IO_U_F_TRIMMED);
					break;
				} else if (io_u->ddir == DDIR_WRITE) {
					io_u->ddir = DDIR_READ;
					io_u->numberio = td->verify_read_issues;
					td->verify_read_issues++;
					populate_verify_io_u(td, io_u);
					break;
				} else {
					put_io_u(td, io_u);
					continue;
				}
			}

			if (!io_u)
				break;
		}

		if (verify_state_should_stop(td, io_u)) {
			put_io_u(td, io_u);
			break;
		}

		if (td->o.verify_async)
			io_u->end_io = verify_io_u_async;
		else
			io_u->end_io = verify_io_u;

		ddir = io_u->ddir;
		if (!td->o.disable_slat)
			fio_gettime(&io_u->start_time, NULL);

		ret = io_u_submit(td, io_u);

		if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
			break;

		/*
		 * if we can queue more, do so. but check if there are
		 * completed io_u's first. 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 || io_in_polling(td))
			ret = wait_for_completions(td, NULL);

		if (ret < 0)
			break;
	}

	check_update_rusage(td);

	if (!td->error) {
		min_events = 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);

	dprint(FD_VERIFY, "exiting loop\n");
}

static bool exceeds_number_ios(struct thread_data *td)
{
	unsigned long long number_ios;

	if (!td->o.number_ios)
		return false;

	number_ios = ddir_rw_sum(td->io_blocks);
	number_ios += td->io_u_queued + td->io_u_in_flight;

	return number_ios >= (td->o.number_ios * td->loops);
}

static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
{
	unsigned long long bytes, limit;

	if (td_rw(td))
		bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
	else if (td_write(td))
		bytes = this_bytes[DDIR_WRITE];
	else if (td_read(td))
		bytes = this_bytes[DDIR_READ];
	else
		bytes = this_bytes[DDIR_TRIM];

	if (td->o.io_size)
		limit = td->o.io_size;
	else
		limit = td->o.size;

	limit *= td->loops;
	return bytes >= limit || exceeds_number_ios(td);
}

static bool io_issue_bytes_exceeded(struct thread_data *td)
{
	return io_bytes_exceeded(td, td->io_issue_bytes);
}

static bool io_complete_bytes_exceeded(struct thread_data *td)
{
	return io_bytes_exceeded(td, td->this_io_bytes);
}

/*
 * 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 bps = td->rate_bps[ddir];

	assert(!(td->flags & TD_F_CHILD));

	if (td->o.rate_process == RATE_PROCESS_POISSON) {
		uint64_t val, iops;

		iops = bps / td->o.min_bs[ddir];
		val = (int64_t) (1000000 / iops) *
				-logf(__rand_0_1(&td->poisson_state[ddir]));
		if (val) {
			dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
					(unsigned long long) 1000000 / val,
					ddir);
		}
		td->last_usec[ddir] += val;
		return td->last_usec[ddir];
	} else if (bps) {
		uint64_t bytes = td->rate_io_issue_bytes[ddir];
		uint64_t secs = bytes / bps;
		uint64_t remainder = bytes % bps;

		return remainder * 1000000 / bps + secs * 1000000;
	}

	return 0;
}

static void init_thinktime(struct thread_data *td)
{
	if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
		td->thinktime_blocks_counter = td->io_blocks;
	else
		td->thinktime_blocks_counter = td->io_issues;
	td->last_thinktime = td->epoch;
	td->last_thinktime_blocks = 0;
}

static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
			     struct timespec *time)
{
	unsigned long long b;
	unsigned long long runtime_left;
	uint64_t total;
	int left;
	struct timespec now;
	bool stall = false;

	if (td->o.thinktime_iotime) {
		fio_gettime(&now, NULL);
		if (utime_since(&td->last_thinktime, &now)
		    >= td->o.thinktime_iotime) {
			stall = true;
		} else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
			/*
			 * When thinktime_iotime is set and thinktime_blocks is
			 * not set, skip the thinktime_blocks check, since
			 * thinktime_blocks default value 1 does not work
			 * together with thinktime_iotime.
			 */
			return;
		}

	}

	b = ddir_rw_sum(td->thinktime_blocks_counter);
	if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
		stall = true;

	if (!stall)
		return;

	io_u_quiesce(td);

	left = td->o.thinktime_spin;
	if (td->o.timeout) {
		runtime_left = td->o.timeout - utime_since_now(&td->epoch);
		if (runtime_left < (unsigned long long)left)
			left = runtime_left;
	}

	total = 0;
	if (left)
		total = usec_spin(left);

	/*
	 * usec_spin() might run for slightly longer than intended in a VM
	 * where the vCPU could get descheduled or the hypervisor could steal
	 * CPU time. Ensure "left" doesn't become negative.
	 */
	if (total < td->o.thinktime)
		left = td->o.thinktime - total;
	else
		left = 0;

	if (td->o.timeout) {
		runtime_left = td->o.timeout - utime_since_now(&td->epoch);
		if (runtime_left < (unsigned long long)left)
			left = runtime_left;
	}

	if (left)
		total += usec_sleep(td, left);

	/*
	 * If we're ignoring thinktime for the rate, add the number of bytes
	 * we would have done while sleeping, minus one block to ensure we
	 * start issuing immediately after the sleep.
	 */
	if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
		uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
		uint64_t bs = td->o.min_bs[ddir];
		uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
		uint64_t over;

		if (usperop <= total)
			over = bs;
		else
			over = (usperop - total) / usperop * -bs;

		td->rate_io_issue_bytes[ddir] += (missed - over);
		/* adjust for rate_process=poisson */
		td->last_usec[ddir] += total;
	}

	if (time && should_check_rate(td))
		fio_gettime(time, NULL);

	td->last_thinktime_blocks = b;
	if (td->o.thinktime_iotime) {
		fio_gettime(&now, NULL);
		td->last_thinktime = now;
	}
}

/*
 * Main IO worker function. It retrieves io_u's to process and queues
 * and reaps them, checking for rate and errors along the way.
 *
 * Returns number of bytes written and trimmed.
 */
static void do_io(struct thread_data *td, uint64_t *bytes_done)
{
	unsigned int i;
	int ret = 0;
	uint64_t total_bytes, bytes_issued = 0;

	for (i = 0; i < DDIR_RWDIR_CNT; i++)
		bytes_done[i] = td->bytes_done[i];

	if (in_ramp_time(td))
		td_set_runstate(td, TD_RAMP);
	else
		td_set_runstate(td, TD_RUNNING);

	lat_target_init(td);

	total_bytes = td->o.size;
	/*
	* Allow random overwrite workloads to write up to io_size
	* before starting verification phase as 'size' doesn't apply.
	*/
	if (td_write(td) && td_random(td) && td->o.norandommap)
		total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
	/*
	 * If verify_backlog is enabled, we'll run the verify in this
	 * handler as well. For that case, we may need up to twice the
	 * amount of bytes.
	 */
	if (td->o.verify != VERIFY_NONE &&
	   (td_write(td) && td->o.verify_backlog))
		total_bytes += td->o.size;

	/* In trimwrite mode, each byte is trimmed and then written, so
	 * allow total_bytes or number of ios to be twice as big */
	if (td_trimwrite(td)) {
		total_bytes += td->total_io_size;
		td->o.number_ios *= 2;
	}

	while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
		(!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
		td->o.time_based) {
		struct timespec comp_time;
		struct io_u *io_u;
		int full;
		enum fio_ddir ddir;

		check_update_rusage(td);

		if (td->terminate || td->done)
			break;

		update_ts_cache(td);

		if (runtime_exceeded(td, &td->ts_cache)) {
			__update_ts_cache(td);
			if (runtime_exceeded(td, &td->ts_cache)) {
				fio_mark_td_terminate(td);
				break;
			}
		}

		if (flow_threshold_exceeded(td))
			continue;

		/*
		 * Break if we exceeded the bytes. The exception is time
		 * based runs, but we still need to break out of the loop
		 * for those to run verification, if enabled.
		 * Jobs read from iolog do not use this stop condition.
		 */
		if (bytes_issued >= total_bytes &&
		    !td->o.read_iolog_file &&
		    (!td->o.time_based ||
		     (td->o.time_based && td->o.verify != VERIFY_NONE)))
			break;

		io_u = get_io_u(td);
		if (IS_ERR_OR_NULL(io_u)) {
			int err = PTR_ERR(io_u);

			io_u = NULL;
			ddir = DDIR_INVAL;
			if (err == -EBUSY) {
				ret = FIO_Q_BUSY;
				goto reap;
			}
			if (td->o.latency_target)
				goto reap;
			break;
		}

		if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) {
			if (!(io_u->flags & IO_U_F_PATTERN_DONE)) {
				io_u_set(td, io_u, IO_U_F_PATTERN_DONE);
				io_u->numberio = td->io_issues[io_u->ddir];
				populate_verify_io_u(td, io_u);
			}
		}

		ddir = io_u->ddir;

		/*
		 * Add verification end_io handler if:
		 *	- Asked to verify (!td_rw(td))
		 *	- Or the io_u is from our verify list (mixed write/ver)
		 */
		if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
		    ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {

			if (verify_state_should_stop(td, io_u)) {
				put_io_u(td, io_u);
				break;
			}

			if (td->o.verify_async)
				io_u->end_io = verify_io_u_async;
			else
				io_u->end_io = verify_io_u;
			td_set_runstate(td, TD_VERIFYING);
		} else if (in_ramp_time(td))
			td_set_runstate(td, TD_RAMP);
		else
			td_set_runstate(td, TD_RUNNING);

		/*
		 * Always log IO before it's issued, so we know the specific
		 * order of it. The logged unit will track when the IO has
		 * completed.
		 */
		if (td_write(td) && io_u->ddir == DDIR_WRITE &&
		    td->o.do_verify &&
		    td->o.verify != VERIFY_NONE &&
		    !td->o.experimental_verify)
			log_io_piece(td, io_u);

		if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
			const unsigned long long blen = io_u->xfer_buflen;
			const enum fio_ddir __ddir = acct_ddir(io_u);

			if (td->error)
				break;

			workqueue_enqueue(&td->io_wq, &io_u->work);
			ret = FIO_Q_QUEUED;

			if (ddir_rw(__ddir)) {
				td->io_issues[__ddir]++;
				td->io_issue_bytes[__ddir] += blen;
				td->rate_io_issue_bytes[__ddir] += blen;
			}

			if (should_check_rate(td)) {
				td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
				fio_gettime(&comp_time, NULL);
			}

		} else {
			ret = io_u_submit(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 || io_in_polling(td))
				ret = wait_for_completions(td, &comp_time);
		}
		if (ret < 0)
			break;

		if (ddir_rw(ddir) && td->o.thinkcycles)
			cycles_spin(td->o.thinkcycles);

		if (ddir_rw(ddir) && td->o.thinktime)
			handle_thinktime(td, ddir, &comp_time);

		if (!ddir_rw_sum(td->bytes_done) &&
		    !td_ioengine_flagged(td, FIO_NOIO))
			continue;

		if (!in_ramp_time(td) && should_check_rate(td)) {
			if (check_min_rate(td, &comp_time)) {
				if (exitall_on_terminate || td->o.exitall_error)
					fio_terminate_threads(td->groupid, td->o.exit_what);
				td_verror(td, EIO, "check_min_rate");
				break;
			}
		}
		if (!in_ramp_time(td) && td->o.latency_target)
			lat_target_check(td);
	}

	check_update_rusage(td);

	if (td->trim_entries)
		log_err("fio: %lu trim entries leaked?\n", td->trim_entries);

	if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
		td->error = 0;
		fio_mark_td_terminate(td);
	}
	if (!td->error) {
		struct fio_file *f;

		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 || td->error == EDQUOT))
				td->error = 0;
		}

		if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
			td_set_runstate(td, TD_FSYNCING);

			for_each_file(td, f, i) {
				if (!fio_file_fsync(td, f))
					continue;

				log_err("fio: end_fsync failed for file %s\n",
								f->file_name);
			}
		}
	} else {
		if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
			workqueue_flush(&td->io_wq);
		cleanup_pending_aio(td);
	}

	/*
	 * stop job if we failed doing any IO
	 */
	if (!ddir_rw_sum(td->this_io_bytes))
		td->done = 1;

	for (i = 0; i < DDIR_RWDIR_CNT; i++)
		bytes_done[i] = td->bytes_done[i] - bytes_done[i];
}

static void free_file_completion_logging(struct thread_data *td)
{
	struct fio_file *f;
	unsigned int i;

	for_each_file(td, f, i) {
		if (!f->last_write_comp)
			break;
		sfree(f->last_write_comp);
	}
}

static int init_file_completion_logging(struct thread_data *td,
					unsigned int depth)
{
	struct fio_file *f;
	unsigned int i;

	if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
		return 0;

	for_each_file(td, f, i) {
		f->last_write_comp = scalloc(depth, sizeof(uint64_t));
		if (!f->last_write_comp)
			goto cleanup;
	}

	return 0;

cleanup:
	free_file_completion_logging(td);
	log_err("fio: failed to alloc write comp data\n");
	return 1;
}

static void cleanup_io_u(struct thread_data *td)
{
	struct io_u *io_u;

	while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {

		if (td->io_ops->io_u_free)
			td->io_ops->io_u_free(td, io_u);

		fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
	}

	free_io_mem(td);

	io_u_rexit(&td->io_u_requeues);
	io_u_qexit(&td->io_u_freelist, false);
	io_u_qexit(&td->io_u_all, td_offload_overlap(td));

	free_file_completion_logging(td);
}

static int init_io_u(struct thread_data *td)
{
	struct io_u *io_u;
	int cl_align, i, max_units;
	int err;

	max_units = td->o.iodepth;

	err = 0;
	err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
	err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
	err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));

	if (err) {
		log_err("fio: failed setting up IO queues\n");
		return 1;
	}

	cl_align = os_cache_line_size();

	for (i = 0; i < max_units; i++) {
		void *ptr;

		if (td->terminate)
			return 1;

		ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
		if (!ptr) {
			log_err("fio: unable to allocate aligned memory\n");
			return 1;
		}

		io_u = ptr;
		memset(io_u, 0, sizeof(*io_u));
		INIT_FLIST_HEAD(&io_u->verify_list);
		dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);

		io_u->index = i;
		io_u->flags = IO_U_F_FREE;
		io_u_qpush(&td->io_u_freelist, io_u);

		/*
		 * io_u never leaves this stack, used for iteration of all
		 * io_u buffers.
		 */
		io_u_qpush(&td->io_u_all, io_u);

		if (td->io_ops->io_u_init) {
			int ret = td->io_ops->io_u_init(td, io_u);

			if (ret) {
				log_err("fio: failed to init engine data: %d\n", ret);
				return 1;
			}
		}
	}

	if (init_io_u_buffers(td))
		return 1;

	if (init_file_completion_logging(td, max_units))
		return 1;

	return 0;
}

int init_io_u_buffers(struct thread_data *td)
{
	struct io_u *io_u;
	unsigned long long max_bs, min_write, trim_bs = 0;
	int i, max_units;
	int data_xfer = 1;
	char *p;

	max_units = td->o.iodepth;
	max_bs = td_max_bs(td);
	min_write = td->o.min_bs[DDIR_WRITE];
	td->orig_buffer_size = (unsigned long long) max_bs
					* (unsigned long long) max_units;

	if (td_trim(td) && td->o.num_range > 1) {
		trim_bs = td->o.num_range * sizeof(struct trim_range);
		td->orig_buffer_size = trim_bs
					* (unsigned long long) max_units;
	}

	/*
	 * For reads, writes, and multi-range trim operations we need a
	 * data buffer
	 */
	if (td_ioengine_flagged(td, FIO_NOIO) ||
	    !(td_read(td) || td_write(td) || (td_trim(td) && td->o.num_range > 1)))
		data_xfer = 0;

	/*
	 * if we may later need to do address alignment, then add any
	 * possible adjustment here so that we don't cause a buffer
	 * overflow later. this adjustment may be too much if we get
	 * lucky and the allocator gives us an aligned address.
	 */
	if (td->o.odirect || td->o.mem_align ||
	    td_ioengine_flagged(td, FIO_RAWIO))
		td->orig_buffer_size += page_mask + td->o.mem_align;

	if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
		unsigned long long bs;

		bs = td->orig_buffer_size + td->o.hugepage_size - 1;
		td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
	}

	if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
		log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
		return 1;
	}

	if (data_xfer && allocate_io_mem(td))
		return 1;

	if (td->o.odirect || td->o.mem_align ||
	    td_ioengine_flagged(td, FIO_RAWIO))
		p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
	else
		p = td->orig_buffer;

	for (i = 0; i < max_units; i++) {
		io_u = td->io_u_all.io_us[i];
		dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);

		if (data_xfer) {
			io_u->buf = p;
			dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);

			if (td_write(td))
				io_u_fill_buffer(td, io_u, min_write, max_bs);
			if (td_write(td) && td->o.verify_pattern_bytes) {
				/*
				 * Fill the buffer with the pattern if we are
				 * going to be doing writes.
				 */
				fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
			}
		}
		if (td_trim(td) && td->o.num_range > 1)
			p += trim_bs;
		else
			p += max_bs;
	}

	return 0;
}

#ifdef FIO_HAVE_IOSCHED_SWITCH
/*
 * These functions are Linux specific.
 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
 */
static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
{
	char tmp[256], tmp2[128], *p;
	FILE *f;
	int ret;

	assert(file->du && file->du->sysfs_root);
	sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);

	f = fopen(tmp, "r+");
	if (!f) {
		if (errno == ENOENT) {
			log_err("fio: os or kernel doesn't support IO scheduler"
				" switching\n");
			return 0;
		}
		td_verror(td, errno, "fopen iosched");
		return 1;
	}

	/*
	 * Set io scheduler.
	 */
	ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
	if (ferror(f) || ret != 1) {
		td_verror(td, errno, "fwrite");
		fclose(f);
		return 1;
	}

	rewind(f);

	/*
	 * Read back and check that the selected scheduler is now the default.
	 */
	ret = fread(tmp, 1, sizeof(tmp) - 1, f);
	if (ferror(f) || ret < 0) {
		td_verror(td, errno, "fread");
		fclose(f);
		return 1;
	}
	tmp[ret] = '\0';
	/*
	 * either a list of io schedulers or "none\n" is expected. Strip the
	 * trailing newline.
	 */
	p = tmp;
	strsep(&p, "\n");

	/*
	 * Write to "none" entry doesn't fail, so check the result here.
	 */
	if (!strcmp(tmp, "none")) {
		log_err("fio: io scheduler is not tunable\n");
		fclose(f);
		return 0;
	}

	sprintf(tmp2, "[%s]", td->o.ioscheduler);
	if (!strstr(tmp, tmp2)) {
		log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
		td_verror(td, EINVAL, "iosched_switch");
		fclose(f);
		return 1;
	}

	fclose(f);
	return 0;
}

static int switch_ioscheduler(struct thread_data *td)
{
	struct fio_file *f;
	unsigned int i;
	int ret = 0;

	if (td_ioengine_flagged(td, FIO_DISKLESSIO))
		return 0;

	assert(td->files && td->files[0]);

	for_each_file(td, f, i) {

		/* Only consider regular files and block device files */
		switch (f->filetype) {
		case FIO_TYPE_FILE:
		case FIO_TYPE_BLOCK:
			/*
			 * Make sure that the device hosting the file could
			 * be determined.
			 */
			if (!f->du)
				continue;
			break;
		case FIO_TYPE_CHAR:
		case FIO_TYPE_PIPE:
		default:
			continue;
		}

		ret = set_ioscheduler(td, f);
		if (ret)
			return ret;
	}

	return 0;
}

#else

static int switch_ioscheduler(struct thread_data *td)
{
	return 0;
}

#endif /* FIO_HAVE_IOSCHED_SWITCH */

static bool keep_running(struct thread_data *td)
{
	unsigned long long limit;

	if (td->done)
		return false;
	if (td->terminate)
		return false;
	if (td->o.time_based)
		return true;
	if (td->o.loops) {
		td->o.loops--;
		return true;
	}
	if (exceeds_number_ios(td))
		return false;

	if (td->o.io_size)
		limit = td->o.io_size;
	else
		limit = td->o.size;

	if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
		uint64_t diff;

		/*
		 * If the difference is less than the maximum IO size, we
		 * are done.
		 */
		diff = limit - ddir_rw_sum(td->io_bytes);
		if (diff < td_max_bs(td))
			return false;

		if (fio_files_done(td) && !td->o.io_size)
			return false;

		return true;
	}

	return false;
}

static int exec_string(struct thread_options *o, const char *string,
		       const char *mode)
{
	int ret;
	char *str;

	if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
		return -1;

	log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
		 o->name, mode);
	ret = system(str);
	if (ret == -1)
		log_err("fio: exec of cmd <%s> failed\n", str);

	free(str);
	return ret;
}

/*
 * Dry run to compute correct state of numberio for verification.
 */
static uint64_t do_dry_run(struct thread_data *td)
{
	td_set_runstate(td, TD_RUNNING);

	while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
		(!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
		struct io_u *io_u;
		int ret;

		if (td->terminate || td->done)
			break;

		io_u = get_io_u(td);
		if (IS_ERR_OR_NULL(io_u))
			break;

		io_u_set(td, io_u, IO_U_F_FLIGHT);
		io_u->error = 0;
		io_u->resid = 0;
		if (ddir_rw(acct_ddir(io_u)))
			td->io_issues[acct_ddir(io_u)]++;
		if (ddir_rw(io_u->ddir)) {
			io_u_mark_depth(td, 1);
			td->ts.total_io_u[io_u->ddir]++;
		}

		if (td_write(td) && io_u->ddir == DDIR_WRITE &&
		    td->o.do_verify &&
		    td->o.verify != VERIFY_NONE &&
		    !td->o.experimental_verify)
			log_io_piece(td, io_u);

		ret = io_u_sync_complete(td, io_u);
		(void) ret;
	}

	return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
}

struct fork_data {
	struct thread_data *td;
	struct sk_out *sk_out;
};

/*
 * 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)
{
	struct fork_data *fd = data;
	unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
	struct thread_data *td = fd->td;
	struct thread_options *o = &td->o;
	struct sk_out *sk_out = fd->sk_out;
	uint64_t bytes_done[DDIR_RWDIR_CNT];
	int deadlock_loop_cnt;
	bool clear_state;
	int ret;

	sk_out_assign(sk_out);
	free(fd);

	if (!o->use_thread) {
		setsid();
		td->pid = getpid();
	} else
		td->pid = gettid();

	fio_local_clock_init();

	dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);

	if (is_backend)
		fio_server_send_start(td);

	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);
	td->io_hist_tree = RB_ROOT;

	ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
	if (ret) {
		td_verror(td, ret, "mutex_cond_init_pshared");
		goto err;
	}
	ret = cond_init_pshared(&td->verify_cond);
	if (ret) {
		td_verror(td, ret, "mutex_cond_pshared");
		goto err;
	}

	td_set_runstate(td, TD_INITIALIZED);
	dprint(FD_MUTEX, "up startup_sem\n");
	fio_sem_up(startup_sem);
	dprint(FD_MUTEX, "wait on td->sem\n");
	fio_sem_down(td->sem);
	dprint(FD_MUTEX, "done waiting on td->sem\n");

	/*
	 * A new gid requires privilege, so we need to do this before setting
	 * the uid.
	 */
	if (o->gid != -1U && setgid(o->gid)) {
		td_verror(td, errno, "setgid");
		goto err;
	}
	if (o->uid != -1U && setuid(o->uid)) {
		td_verror(td, errno, "setuid");
		goto err;
	}

	td_zone_gen_index(td);

	/*
	 * Do this early, we don't want the compress threads to be limited
	 * to the same CPUs as the IO workers. So do this before we set
	 * any potential CPU affinity
	 */
	if (iolog_compress_init(td, sk_out))
		goto err;

	/*
	 * If we have a gettimeofday() thread, make sure we exclude that
	 * thread from this job
	 */
	if (o->gtod_cpu)
		fio_cpu_clear(&o->cpumask, o->gtod_cpu);

	/*
	 * Set affinity first, in case it has an impact on the memory
	 * allocations.
	 */
	if (fio_option_is_set(o, cpumask)) {
		if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
			ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
			if (!ret) {
				log_err("fio: no CPUs set\n");
				log_err("fio: Try increasing number of available CPUs\n");
				td_verror(td, EINVAL, "cpus_split");
				goto err;
			}
		}
		ret = fio_setaffinity(td->pid, o->cpumask);
		if (ret == -1) {
			td_verror(td, errno, "cpu_set_affinity");
			goto err;
		}
	}

#ifdef CONFIG_LIBNUMA
	/* numa node setup */
	if (fio_option_is_set(o, numa_cpunodes) ||
	    fio_option_is_set(o, numa_memnodes)) {
		struct bitmask *mask;

		if (numa_available() < 0) {
			td_verror(td, errno, "Does not support NUMA API\n");
			goto err;
		}

		if (fio_option_is_set(o, numa_cpunodes)) {
			mask = numa_parse_nodestring(o->numa_cpunodes);
			ret = numa_run_on_node_mask(mask);
			numa_free_nodemask(mask);
			if (ret == -1) {
				td_verror(td, errno, \
					"numa_run_on_node_mask failed\n");
				goto err;
			}
		}

		if (fio_option_is_set(o, numa_memnodes)) {
			mask = NULL;
			if (o->numa_memnodes)
				mask = numa_parse_nodestring(o->numa_memnodes);

			switch (o->numa_mem_mode) {
			case MPOL_INTERLEAVE:
				numa_set_interleave_mask(mask);
				break;
			case MPOL_BIND:
				numa_set_membind(mask);
				break;
			case MPOL_LOCAL:
				numa_set_localalloc();
				break;
			case MPOL_PREFERRED:
				numa_set_preferred(o->numa_mem_prefer_node);
				break;
			case MPOL_DEFAULT:
			default:
				break;
			}

			if (mask)
				numa_free_nodemask(mask);

		}
	}
#endif

	if (fio_pin_memory(td))
		goto err;

	/*
	 * May alter parameters that init_io_u() will use, so we need to
	 * do this first.
	 */
	if (!init_iolog(td))
		goto err;

	/* ioprio_set() has to be done before td_io_init() */
	if (fio_option_is_set(o, ioprio) ||
	    fio_option_is_set(o, ioprio_class) ||
	    fio_option_is_set(o, ioprio_hint)) {
		ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class,
				 o->ioprio, o->ioprio_hint);
		if (ret == -1) {
			td_verror(td, errno, "ioprio_set");
			goto err;
		}
		td->ioprio = ioprio_value(o->ioprio_class, o->ioprio,
					  o->ioprio_hint);
		td->ts.ioprio = td->ioprio;
	}

	if (td_io_init(td))
		goto err;

	if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1 && td->o.io_submit_mode != IO_MODE_OFFLOAD) {
		log_info("note: both iodepth >= 1 and synchronous I/O engine "
			 "are selected, queue depth will be capped at 1\n");
	}

	if (init_io_u(td))
		goto err;

	if (td->io_ops->post_init && td->io_ops->post_init(td))
		goto err;

	if (o->verify_async && verify_async_init(td))
		goto err;

	if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
		goto err;

	errno = 0;
	if (nice(o->nice) == -1 && errno != 0) {
		td_verror(td, errno, "nice");
		goto err;
	}

	if (o->ioscheduler && switch_ioscheduler(td))
		goto err;

	if (!o->create_serialize && setup_files(td))
		goto err;

	if (!init_random_map(td))
		goto err;

	if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
		goto err;

	if (o->pre_read && !pre_read_files(td))
		goto err;

	fio_verify_init(td);

	if (rate_submit_init(td, sk_out))
		goto err;

	set_epoch_time(td, o->log_alternate_epoch_clock_id, o->job_start_clock_id);
	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));
	memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));

	init_thinktime(td);

	if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
			o->ratemin[DDIR_TRIM]) {
	        memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
					sizeof(td->bw_sample_time));
	        memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
					sizeof(td->bw_sample_time));
	        memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
					sizeof(td->bw_sample_time));
	}

	memset(bytes_done, 0, sizeof(bytes_done));
	clear_state = false;

	while (keep_running(td)) {
		uint64_t verify_bytes;

		fio_gettime(&td->start, NULL);
		memcpy(&td->ts_cache, &td->start, sizeof(td->start));

		if (clear_state) {
			clear_io_state(td, 0);

			if (o->unlink_each_loop && unlink_all_files(td))
				break;
		}

		prune_io_piece_log(td);

		if (td->o.verify_only && td_write(td))
			verify_bytes = do_dry_run(td);
		else {
			if (!td->o.rand_repeatable)
				/* save verify rand state to replay hdr seeds later at verify */
				frand_copy(&td->verify_state_last_do_io, &td->verify_state);
			do_io(td, bytes_done);
			if (!td->o.rand_repeatable)
				frand_copy(&td->verify_state, &td->verify_state_last_do_io);
			if (!ddir_rw_sum(bytes_done)) {
				fio_mark_td_terminate(td);
				verify_bytes = 0;
			} else {
				verify_bytes = bytes_done[DDIR_WRITE] +
						bytes_done[DDIR_TRIM];
			}
		}

		/*
		 * If we took too long to shut down, the main thread could
		 * already consider us reaped/exited. If that happens, break
		 * out and clean up.
		 */
		if (td->runstate >= TD_EXITED)
			break;

		clear_state = true;

		/*
		 * Make sure we've successfully updated the rusage stats
		 * before waiting on the stat mutex. Otherwise we could have
		 * the stat thread holding stat mutex and waiting for
		 * the rusage_sem, which would never get upped because
		 * this thread is waiting for the stat mutex.
		 */
		deadlock_loop_cnt = 0;
		do {
			check_update_rusage(td);
			if (!fio_sem_down_trylock(stat_sem))
				break;
			usleep(1000);
			if (deadlock_loop_cnt++ > 5000) {
				log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
				td->error = EDEADLK;
				goto err;
			}
		} while (1);

		if (td->io_bytes[DDIR_READ] && (td_read(td) ||
			((td->flags & TD_F_VER_BACKLOG) && td_write(td))))
			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_sem_up(stat_sem);

		if (td->error || td->terminate)
			break;

		if (!o->do_verify ||
		    o->verify == VERIFY_NONE ||
		    td_ioengine_flagged(td, FIO_UNIDIR))
			continue;

		clear_io_state(td, 0);

		fio_gettime(&td->start, NULL);

		do_verify(td, verify_bytes);

		/*
		 * See comment further up for why this is done here.
		 */
		check_update_rusage(td);

		fio_sem_down(stat_sem);
		update_runtime(td, elapsed_us, DDIR_READ);
		fio_gettime(&td->start, NULL);
		fio_sem_up(stat_sem);

		if (td->error || td->terminate)
			break;
	}

	/*
	 * Acquire this lock if we were doing overlap checking in
	 * offload mode so that we don't clean up this job while
	 * another thread is checking its io_u's for overlap
	 */
	if (td_offload_overlap(td)) {
		int res;

		res = pthread_mutex_lock(&overlap_check);
		if (res) {
			td->error = errno;
			goto err;
		}
	}
	td_set_runstate(td, TD_FINISHING);
	if (td_offload_overlap(td)) {
		int res;

		res = pthread_mutex_unlock(&overlap_check);
		if (res) {
			td->error = errno;
			goto err;
		}
	}

	update_rusage_stat(td);
	td->ts.total_run_time = mtime_since_now(&td->epoch);
	for_each_rw_ddir(ddir) {
		td->ts.io_bytes[ddir] = td->io_bytes[ddir];
	}

	if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
	    (td->o.verify != VERIFY_NONE && td_write(td)))
		verify_save_state(td->thread_number);

	fio_unpin_memory(td);

	td_writeout_logs(td, true);

	iolog_compress_exit(td);
	rate_submit_exit(td);

	if (o->exec_postrun)
		exec_string(o, o->exec_postrun, "postrun");

	if (exitall_on_terminate || (o->exitall_error && td->error))
		fio_terminate_threads(td->groupid, td->o.exit_what);

err:
	if (td->error)
		log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
							td->verror);

	if (o->verify_async)
		verify_async_exit(td);

	close_and_free_files(td);
	cleanup_io_u(td);
	close_ioengine(td);
	cgroup_shutdown(td, cgroup_mnt);
	verify_free_state(td);
	td_zone_free_index(td);

	if (fio_option_is_set(o, cpumask)) {
		ret = fio_cpuset_exit(&o->cpumask);
		if (ret)
			td_verror(td, ret, "fio_cpuset_exit");
	}

	/*
	 * do this very late, it will log file closing as well
	 */
	if (o->write_iolog_file)
		write_iolog_close(td);
	if (td->io_log_rfile)
		fclose(td->io_log_rfile);

	td_set_runstate(td, TD_EXITED);

	/*
	 * Do this last after setting our runstate to exited, so we
	 * know that the stat thread is signaled.
	 */
	check_update_rusage(td);

	sk_out_drop();
	return (void *) (uintptr_t) td->error;
}

/*
 * Run over the job map and reap the threads that have exited, if any.
 */
static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
			 uint64_t *m_rate)
{
	unsigned int cputhreads, realthreads, pending;
	int status, ret;

	/*
	 * reap exited threads (TD_EXITED -> TD_REAPED)
	 */
	realthreads = pending = cputhreads = 0;
	for_each_td(td) {
		int flags = 0;

		if (!strcmp(td->o.ioengine, "cpuio"))
			cputhreads++;
		else
			realthreads++;

		if (!td->pid) {
			pending++;
			continue;
		}
		if (td->runstate == TD_REAPED)
			continue;
		if (td->o.use_thread) {
			if (td->runstate == TD_EXITED) {
				td_set_runstate(td, TD_REAPED);
				goto reaped;
			}
			continue;
		}

		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",
						(int) td->pid, td->runstate);
				td->sig = ECHILD;
				td_set_runstate(td, TD_REAPED);
				goto reaped;
			}
			perror("waitpid");
		} else if (ret == td->pid) {
			if (WIFSIGNALED(status)) {
				int sig = WTERMSIG(status);

				if (sig != SIGTERM && sig != SIGUSR2)
					log_err("fio: pid=%d, got signal=%d\n",
							(int) td->pid, sig);
				td->sig = 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 the job is stuck, do a forceful timeout of it and
		 * move on.
		 */
		if (td->terminate &&
		    td->runstate < TD_FSYNCING &&
		    time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
			log_err("fio: job '%s' (state=%d) hasn't exited in "
				"%lu seconds, it appears to be stuck. Doing "
				"forceful exit of this job.\n",
				td->o.name, td->runstate,
				(unsigned long) time_since_now(&td->terminate_time));
			td_set_runstate(td, TD_REAPED);
			goto reaped;
		}

		/*
		 * thread is not dead, continue
		 */
		pending++;
		continue;
reaped:
		(*nr_running)--;
		(*m_rate) -= ddir_rw_sum(td->o.ratemin);
		(*t_rate) -= ddir_rw_sum(td->o.rate);
		if (!td->pid)
			pending--;

		if (td->error)
			exit_value++;

		done_secs += mtime_since_now(&td->epoch) / 1000;
		profile_td_exit(td);
		flow_exit_job(td);
	} end_for_each();

	if (*nr_running == cputhreads && !pending && realthreads)
		fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
}

static bool __check_trigger_file(void)
{
	struct stat sb;

	if (!trigger_file)
		return false;

	if (stat(trigger_file, &sb))
		return false;

	if (unlink(trigger_file) < 0)
		log_err("fio: failed to unlink %s: %s\n", trigger_file,
							strerror(errno));

	return true;
}

static bool trigger_timedout(void)
{
	if (trigger_timeout)
		if (time_since_genesis() >= trigger_timeout) {
			trigger_timeout = 0;
			return true;
		}

	return false;
}

void exec_trigger(const char *cmd)
{
	int ret;

	if (!cmd || cmd[0] == '\0')
		return;

	ret = system(cmd);
	if (ret == -1)
		log_err("fio: failed executing %s trigger\n", cmd);
}

void check_trigger_file(void)
{
	if (__check_trigger_file() || trigger_timedout()) {
		if (nr_clients)
			fio_clients_send_trigger(trigger_remote_cmd);
		else {
			verify_save_state(IO_LIST_ALL);
			fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
			exec_trigger(trigger_cmd);
		}
	}
}

static int fio_verify_load_state(struct thread_data *td)
{
	int ret;

	if (!td->o.verify_state)
		return 0;

	if (is_backend) {
		void *data;

		ret = fio_server_get_verify_state(td->o.name,
					td->thread_number - 1, &data);
		if (!ret)
			verify_assign_state(td, data);
	} else {
		char prefix[PATH_MAX];

		if (aux_path)
			sprintf(prefix, "%s%clocal", aux_path,
					FIO_OS_PATH_SEPARATOR);
		else
			strcpy(prefix, "local");
		ret = verify_load_state(td, prefix);
	}

	return ret;
}

static void do_usleep(unsigned int usecs)
{
	check_for_running_stats();
	check_trigger_file();
	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;

	/*
	 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
	 * are mkfs'd and mounted.
	 */
	for_each_file(td, f, i) {
#ifdef FIO_HAVE_CHARDEV_SIZE
		if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
#else
		if (f->filetype != FIO_TYPE_BLOCK)
#endif
			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.\n", f->file_name);
	return true;
}

static bool waitee_running(struct thread_data *me)
{
	const char *waitee = me->o.wait_for;
	const char *self = me->o.name;

	if (!waitee)
		return false;

	for_each_td(td) {
		if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
			continue;

		if (td->runstate < TD_EXITED) {
			dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
					self, td->o.name,
					runstate_to_name(td->runstate));
			return true;
		}
	} end_for_each();

	dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
	return false;
}

/*
 * Main function for kicking off and reaping jobs, as needed.
 */
static void run_threads(struct sk_out *sk_out)
{
	struct thread_data *td;
	unsigned int i, todo, nr_running, nr_started;
	uint64_t m_rate, t_rate;
	uint64_t spent;

	if (fio_gtod_offload && fio_start_gtod_thread())
		return;

	fio_idle_prof_init();

	set_sig_handlers();

	nr_thread = nr_process = 0;
	for_each_td(td) {
		if (check_mount_writes(td))
			return;
		if (td->o.use_thread)
			nr_thread++;
		else
			nr_process++;
	} end_for_each();

	if (output_format & FIO_OUTPUT_NORMAL) {
		struct buf_output out;

		buf_output_init(&out);
		__log_buf(&out, "Starting ");
		if (nr_thread)
			__log_buf(&out, "%d thread%s", nr_thread,
						nr_thread > 1 ? "s" : "");
		if (nr_process) {
			if (nr_thread)
				__log_buf(&out, " and ");
			__log_buf(&out, "%d process%s", nr_process,
						nr_process > 1 ? "es" : "");
		}
		__log_buf(&out, "\n");
		log_info_buf(out.buf, out.buflen);
		buf_output_free(&out);
	}

	todo = thread_number;
	nr_running = 0;
	nr_started = 0;
	m_rate = t_rate = 0;

	for_each_td(td) {
		print_status_init(td->thread_number - 1);

		if (!td->o.create_serialize)
			continue;

		if (fio_verify_load_state(td))
			goto reap;

		/*
		 * do file setup here so it happens sequentially,
		 * we don't want X number of threads getting their
		 * client data interspersed on disk
		 */
		if (setup_files(td)) {
reap:
			exit_value++;
			if (td->error)
				log_err("fio: pid=%d, err=%d/%s\n",
					(int) td->pid, td->error, td->verror);
			td_set_runstate(td, TD_REAPED);
			todo--;
		} else {
			struct fio_file *f;
			unsigned int j;

			/*
			 * for sharing to work, each job must always open
			 * its own files. so close them, if we opened them
			 * for creation
			 */
			for_each_file(td, f, j) {
				if (fio_file_open(f))
					td_io_close_file(td, f);
			}
		}
	} end_for_each();

	/* start idle threads before io threads start to run */
	fio_idle_prof_start();

	set_genesis_time();

	while (todo) {
		struct thread_data *map[REAL_MAX_JOBS];
		struct timespec this_start;
		int this_jobs = 0, left;
		struct fork_data *fd;

		/*
		 * create threads (TD_NOT_CREATED -> TD_CREATED)
		 */
		for_each_td(td) {
			if (td->runstate != TD_NOT_CREATED)
				continue;

			/*
			 * never got a chance to start, killed by other
			 * thread for some reason
			 */
			if (td->terminate) {
				todo--;
				continue;
			}

			if (td->o.start_delay) {
				spent = utime_since_genesis();

				if (td->o.start_delay > spent)
					continue;
			}

			if (td->o.stonewall && (nr_started || nr_running)) {
				dprint(FD_PROCESS, "%s: stonewall wait\n",
							td->o.name);
				break;
			}

			if (waitee_running(td)) {
				dprint(FD_PROCESS, "%s: waiting for %s\n",
						td->o.name, td->o.wait_for);
				continue;
			}

			init_disk_util(td);

			td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
			td->update_rusage = 0;

			/*
			 * Set state to created. Thread will transition
			 * to TD_INITIALIZED when it's done setting up.
			 */
			td_set_runstate(td, TD_CREATED);
			map[this_jobs++] = td;
			nr_started++;

			fd = calloc(1, sizeof(*fd));
			fd->td = td;
			fd->sk_out = sk_out;

			if (td->o.use_thread) {
				int ret;

				dprint(FD_PROCESS, "will pthread_create\n");
				ret = pthread_create(&td->thread, NULL,
							thread_main, fd);
				if (ret) {
					log_err("pthread_create: %s\n",
							strerror(ret));
					free(fd);
					nr_started--;
					break;
				}
				fd = NULL;
				ret = pthread_detach(td->thread);
				if (ret)
					log_err("pthread_detach: %s",
							strerror(ret));
			} else {
				pid_t pid;
				void *eo;
				dprint(FD_PROCESS, "will fork\n");
				eo = td->eo;
				read_barrier();
				pid = fork();
				if (!pid) {
					int ret;

					ret = (int)(uintptr_t)thread_main(fd);
					_exit(ret);
				} else if (__td_index == fio_debug_jobno)
					*fio_debug_jobp = pid;
				free(eo);
				free(fd);
				fd = NULL;
			}
			dprint(FD_MUTEX, "wait on startup_sem\n");
			if (fio_sem_down_timeout(startup_sem, 10000)) {
				log_err("fio: job startup hung? exiting.\n");
				fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
				fio_abort = true;
				nr_started--;
				free(fd);
				break;
			}
			dprint(FD_MUTEX, "done waiting on startup_sem\n");
		} end_for_each();

		/*
		 * Wait for the started threads to transition to
		 * TD_INITIALIZED.
		 */
		fio_gettime(&this_start, NULL);
		left = this_jobs;
		while (left && !fio_abort) {
			if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
				break;

			do_usleep(100000);

			for (i = 0; i < this_jobs; i++) {
				td = map[i];
				if (!td)
					continue;
				if (td->runstate == TD_INITIALIZED) {
					map[i] = NULL;
					left--;
				} else if (td->runstate >= TD_EXITED) {
					map[i] = NULL;
					left--;
					todo--;
					nr_running++; /* work-around... */
				}
			}
		}

		if (left) {
			log_err("fio: %d job%s failed to start\n", left,
					left > 1 ? "s" : "");
			for (i = 0; i < this_jobs; i++) {
				td = map[i];
				if (!td)
					continue;
				kill(td->pid, SIGTERM);
			}
			break;
		}

		/*
		 * start created threads (TD_INITIALIZED -> TD_RUNNING).
		 */
		for_each_td(td) {
			if (td->runstate != TD_INITIALIZED)
				continue;

			if (in_ramp_time(td))
				td_set_runstate(td, TD_RAMP);
			else
				td_set_runstate(td, TD_RUNNING);
			nr_running++;
			nr_started--;
			m_rate += ddir_rw_sum(td->o.ratemin);
			t_rate += ddir_rw_sum(td->o.rate);
			todo--;
			fio_sem_up(td->sem);
		} end_for_each();

		reap_threads(&nr_running, &t_rate, &m_rate);

		if (todo)
			do_usleep(100000);
	}

	while (nr_running) {
		reap_threads(&nr_running, &t_rate, &m_rate);
		do_usleep(10000);
	}

	fio_idle_prof_stop();

	update_io_ticks();
}

static void free_disk_util(void)
{
	disk_util_prune_entries();
	helper_thread_destroy();
}

int fio_backend(struct sk_out *sk_out)
{
	int i;
	if (exec_profile) {
		if (load_profile(exec_profile))
			return 1;
		free(exec_profile);
		exec_profile = NULL;
	}
	if (!thread_number)
		return 0;

	if (write_bw_log) {
		struct log_params p = {
			.log_type = IO_LOG_TYPE_BW,
		};

		setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
		setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
		setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
	}

	if (init_global_dedupe_working_set_seeds()) {
		log_err("fio: failed to initialize global dedupe working set\n");
		return 1;
	}

	startup_sem = fio_sem_init(FIO_SEM_LOCKED);
	if (!sk_out)
		is_local_backend = true;
	if (startup_sem == NULL)
		return 1;

	set_genesis_time();
	stat_init();
	if (helper_thread_create(startup_sem, sk_out))
		log_err("fio: failed to create helper thread\n");

	cgroup_list = smalloc(sizeof(*cgroup_list));
	if (cgroup_list)
		INIT_FLIST_HEAD(cgroup_list);

	run_threads(sk_out);

	helper_thread_exit();

	if (!fio_abort) {
		__show_run_stats();
		if (write_bw_log) {
			for (i = 0; i < DDIR_RWDIR_CNT; i++) {
				struct io_log *log = agg_io_log[i];

				flush_log(log, false);
				free_log(log);
			}
		}
	}

	for_each_td(td) {
		struct thread_stat *ts = &td->ts;

		free_clat_prio_stats(ts);
		steadystate_free(td);
		fio_options_free(td);
		fio_dump_options_free(td);
		if (td->rusage_sem) {
			fio_sem_remove(td->rusage_sem);
			td->rusage_sem = NULL;
		}
		fio_sem_remove(td->sem);
		td->sem = NULL;
	} end_for_each();

	free_disk_util();
	if (cgroup_list) {
		cgroup_kill(cgroup_list);
		sfree(cgroup_list);
	}

	fio_sem_remove(startup_sem);
	stat_exit();
	return exit_value;
}