init.c revision 6cfe9a8c89c2b8fb0b697e2f5f852ce6a5e82a2a
1/*
2 * This file contains job initialization and setup functions.
3 */
4#include <stdio.h>
5#include <stdlib.h>
6#include <unistd.h>
7#include <fcntl.h>
8#include <ctype.h>
9#include <string.h>
10#include <errno.h>
11#include <sys/ipc.h>
12#include <sys/types.h>
13#include <sys/stat.h>
14
15#include "fio.h"
16#ifndef FIO_NO_HAVE_SHM_H
17#include <sys/shm.h>
18#endif
19
20#include "parse.h"
21#include "smalloc.h"
22#include "filehash.h"
23#include "verify.h"
24#include "profile.h"
25#include "server.h"
26#include "idletime.h"
27#include "filelock.h"
28
29#include "lib/getopt.h"
30#include "lib/strcasestr.h"
31
32#include "crc/test.h"
33
34const char fio_version_string[] = FIO_VERSION;
35
36#define FIO_RANDSEED		(0xb1899bedUL)
37
38static char **ini_file;
39static int max_jobs = FIO_MAX_JOBS;
40static int dump_cmdline;
41static long long def_timeout;
42static int parse_only;
43
44static struct thread_data def_thread;
45struct thread_data *threads = NULL;
46static char **job_sections;
47static int nr_job_sections;
48
49int exitall_on_terminate = 0;
50int output_format = FIO_OUTPUT_NORMAL;
51int append_terse_output = 0;
52int eta_print = FIO_ETA_AUTO;
53int eta_new_line = 0;
54FILE *f_out = NULL;
55FILE *f_err = NULL;
56char *exec_profile = NULL;
57int warnings_fatal = 0;
58int terse_version = 3;
59int is_backend = 0;
60int nr_clients = 0;
61int log_syslog = 0;
62
63int write_bw_log = 0;
64int read_only = 0;
65int status_interval = 0;
66
67static int write_lat_log;
68
69static int prev_group_jobs;
70
71unsigned long fio_debug = 0;
72unsigned int fio_debug_jobno = -1;
73unsigned int *fio_debug_jobp = NULL;
74
75static char cmd_optstr[256];
76static int did_arg;
77
78#define FIO_CLIENT_FLAG		(1 << 16)
79
80/*
81 * Command line options. These will contain the above, plus a few
82 * extra that only pertain to fio itself and not jobs.
83 */
84static struct option l_opts[FIO_NR_OPTIONS] = {
85	{
86		.name		= (char *) "output",
87		.has_arg	= required_argument,
88		.val		= 'o' | FIO_CLIENT_FLAG,
89	},
90	{
91		.name		= (char *) "timeout",
92		.has_arg	= required_argument,
93		.val		= 't' | FIO_CLIENT_FLAG,
94	},
95	{
96		.name		= (char *) "latency-log",
97		.has_arg	= required_argument,
98		.val		= 'l' | FIO_CLIENT_FLAG,
99	},
100	{
101		.name		= (char *) "bandwidth-log",
102		.has_arg	= required_argument,
103		.val		= 'b' | FIO_CLIENT_FLAG,
104	},
105	{
106		.name		= (char *) "minimal",
107		.has_arg	= no_argument,
108		.val		= 'm' | FIO_CLIENT_FLAG,
109	},
110	{
111		.name		= (char *) "output-format",
112		.has_arg	= optional_argument,
113		.val		= 'F' | FIO_CLIENT_FLAG,
114	},
115	{
116		.name		= (char *) "append-terse",
117		.has_arg	= optional_argument,
118		.val		= 'f',
119	},
120	{
121		.name		= (char *) "version",
122		.has_arg	= no_argument,
123		.val		= 'v' | FIO_CLIENT_FLAG,
124	},
125	{
126		.name		= (char *) "help",
127		.has_arg	= no_argument,
128		.val		= 'h' | FIO_CLIENT_FLAG,
129	},
130	{
131		.name		= (char *) "cmdhelp",
132		.has_arg	= optional_argument,
133		.val		= 'c' | FIO_CLIENT_FLAG,
134	},
135	{
136		.name		= (char *) "enghelp",
137		.has_arg	= optional_argument,
138		.val		= 'i' | FIO_CLIENT_FLAG,
139	},
140	{
141		.name		= (char *) "showcmd",
142		.has_arg	= no_argument,
143		.val		= 's' | FIO_CLIENT_FLAG,
144	},
145	{
146		.name		= (char *) "readonly",
147		.has_arg	= no_argument,
148		.val		= 'r' | FIO_CLIENT_FLAG,
149	},
150	{
151		.name		= (char *) "eta",
152		.has_arg	= required_argument,
153		.val		= 'e' | FIO_CLIENT_FLAG,
154	},
155	{
156		.name		= (char *) "eta-newline",
157		.has_arg	= required_argument,
158		.val		= 'E' | FIO_CLIENT_FLAG,
159	},
160	{
161		.name		= (char *) "debug",
162		.has_arg	= required_argument,
163		.val		= 'd' | FIO_CLIENT_FLAG,
164	},
165	{
166		.name		= (char *) "parse-only",
167		.has_arg	= no_argument,
168		.val		= 'P' | FIO_CLIENT_FLAG,
169	},
170	{
171		.name		= (char *) "section",
172		.has_arg	= required_argument,
173		.val		= 'x' | FIO_CLIENT_FLAG,
174	},
175	{
176		.name		= (char *) "alloc-size",
177		.has_arg	= required_argument,
178		.val		= 'a' | FIO_CLIENT_FLAG,
179	},
180	{
181		.name		= (char *) "profile",
182		.has_arg	= required_argument,
183		.val		= 'p' | FIO_CLIENT_FLAG,
184	},
185	{
186		.name		= (char *) "warnings-fatal",
187		.has_arg	= no_argument,
188		.val		= 'w' | FIO_CLIENT_FLAG,
189	},
190	{
191		.name		= (char *) "max-jobs",
192		.has_arg	= required_argument,
193		.val		= 'j' | FIO_CLIENT_FLAG,
194	},
195	{
196		.name		= (char *) "terse-version",
197		.has_arg	= required_argument,
198		.val		= 'V' | FIO_CLIENT_FLAG,
199	},
200	{
201		.name		= (char *) "server",
202		.has_arg	= optional_argument,
203		.val		= 'S',
204	},
205	{	.name		= (char *) "daemonize",
206		.has_arg	= required_argument,
207		.val		= 'D',
208	},
209	{
210		.name		= (char *) "client",
211		.has_arg	= required_argument,
212		.val		= 'C',
213	},
214	{
215		.name		= (char *) "cpuclock-test",
216		.has_arg	= no_argument,
217		.val		= 'T',
218	},
219	{
220		.name		= (char *) "crctest",
221		.has_arg	= optional_argument,
222		.val		= 'G',
223	},
224	{
225		.name		= (char *) "idle-prof",
226		.has_arg	= required_argument,
227		.val		= 'I',
228	},
229	{
230		.name		= (char *) "status-interval",
231		.has_arg	= required_argument,
232		.val		= 'L',
233	},
234	{
235		.name		= NULL,
236	},
237};
238
239void free_threads_shm(void)
240{
241	struct shmid_ds sbuf;
242
243	if (threads) {
244		void *tp = threads;
245
246		threads = NULL;
247		shmdt(tp);
248		shmctl(shm_id, IPC_RMID, &sbuf);
249		shm_id = -1;
250	}
251}
252
253static void free_shm(void)
254{
255	if (threads) {
256		file_hash_exit();
257		flow_exit();
258		fio_debug_jobp = NULL;
259		free_threads_shm();
260	}
261
262	options_free(fio_options, &def_thread);
263	fio_filelock_exit();
264	scleanup();
265}
266
267/*
268 * The thread area is shared between the main process and the job
269 * threads/processes. So setup a shared memory segment that will hold
270 * all the job info. We use the end of the region for keeping track of
271 * open files across jobs, for file sharing.
272 */
273static int setup_thread_area(void)
274{
275	void *hash;
276
277	if (threads)
278		return 0;
279
280	/*
281	 * 1024 is too much on some machines, scale max_jobs if
282	 * we get a failure that looks like too large a shm segment
283	 */
284	do {
285		size_t size = max_jobs * sizeof(struct thread_data);
286
287		size += file_hash_size;
288		size += sizeof(unsigned int);
289
290		shm_id = shmget(0, size, IPC_CREAT | 0600);
291		if (shm_id != -1)
292			break;
293		if (errno != EINVAL && errno != ENOMEM && errno != ENOSPC) {
294			perror("shmget");
295			break;
296		}
297
298		max_jobs >>= 1;
299	} while (max_jobs);
300
301	if (shm_id == -1)
302		return 1;
303
304	threads = shmat(shm_id, NULL, 0);
305	if (threads == (void *) -1) {
306		perror("shmat");
307		return 1;
308	}
309
310	memset(threads, 0, max_jobs * sizeof(struct thread_data));
311	hash = (void *) threads + max_jobs * sizeof(struct thread_data);
312	fio_debug_jobp = (void *) hash + file_hash_size;
313	*fio_debug_jobp = -1;
314	file_hash_init(hash);
315
316	flow_init();
317
318	return 0;
319}
320
321static void set_cmd_options(struct thread_data *td)
322{
323	struct thread_options *o = &td->o;
324
325	if (!o->timeout)
326		o->timeout = def_timeout;
327}
328
329/*
330 * Return a free job structure.
331 */
332static struct thread_data *get_new_job(int global, struct thread_data *parent,
333				       int preserve_eo)
334{
335	struct thread_data *td;
336
337	if (global) {
338		set_cmd_options(&def_thread);
339		return &def_thread;
340	}
341	if (setup_thread_area()) {
342		log_err("error: failed to setup shm segment\n");
343		return NULL;
344	}
345	if (thread_number >= max_jobs) {
346		log_err("error: maximum number of jobs (%d) reached.\n",
347				max_jobs);
348		return NULL;
349	}
350
351	td = &threads[thread_number++];
352	*td = *parent;
353
354	td->io_ops = NULL;
355	if (!preserve_eo)
356		td->eo = NULL;
357
358	td->o.uid = td->o.gid = -1U;
359
360	dup_files(td, parent);
361	fio_options_mem_dupe(td);
362
363	profile_add_hooks(td);
364
365	td->thread_number = thread_number;
366
367	if (!parent->o.group_reporting)
368		stat_number++;
369
370	set_cmd_options(td);
371	return td;
372}
373
374static void put_job(struct thread_data *td)
375{
376	if (td == &def_thread)
377		return;
378
379	profile_td_exit(td);
380	flow_exit_job(td);
381
382	if (td->error)
383		log_info("fio: %s\n", td->verror);
384
385	fio_options_free(td);
386	if (td->io_ops)
387		free_ioengine(td);
388
389	memset(&threads[td->thread_number - 1], 0, sizeof(*td));
390	thread_number--;
391}
392
393static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
394{
395	unsigned int bs = td->o.min_bs[ddir];
396
397	assert(ddir_rw(ddir));
398
399	if (td->o.rate[ddir])
400		td->rate_bps[ddir] = td->o.rate[ddir];
401	else
402		td->rate_bps[ddir] = td->o.rate_iops[ddir] * bs;
403
404	if (!td->rate_bps[ddir]) {
405		log_err("rate lower than supported\n");
406		return -1;
407	}
408
409	td->rate_pending_usleep[ddir] = 0;
410	return 0;
411}
412
413static int setup_rate(struct thread_data *td)
414{
415	int ret = 0;
416
417	if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
418		ret = __setup_rate(td, DDIR_READ);
419	if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
420		ret |= __setup_rate(td, DDIR_WRITE);
421	if (td->o.rate[DDIR_TRIM] || td->o.rate_iops[DDIR_TRIM])
422		ret |= __setup_rate(td, DDIR_TRIM);
423
424	return ret;
425}
426
427static int fixed_block_size(struct thread_options *o)
428{
429	return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
430		o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
431		o->min_bs[DDIR_TRIM] == o->max_bs[DDIR_TRIM] &&
432		o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE] &&
433		o->min_bs[DDIR_READ] == o->min_bs[DDIR_TRIM];
434}
435
436
437static unsigned long long get_rand_start_delay(struct thread_data *td)
438{
439	unsigned long long delayrange;
440	unsigned long r;
441
442	delayrange = td->o.start_delay_high - td->o.start_delay;
443
444	if (td->o.use_os_rand) {
445		r = os_random_long(&td->delay_state);
446		delayrange = (unsigned long long) ((double) delayrange * (r / (OS_RAND_MAX + 1.0)));
447	} else {
448		r = __rand(&td->__delay_state);
449		delayrange = (unsigned long long) ((double) delayrange * (r / (FRAND_MAX + 1.0)));
450	}
451
452	delayrange += td->o.start_delay;
453	return delayrange;
454}
455
456/*
457 * Lazy way of fixing up options that depend on each other. We could also
458 * define option callback handlers, but this is easier.
459 */
460static int fixup_options(struct thread_data *td)
461{
462	struct thread_options *o = &td->o;
463	int ret = 0;
464
465#ifndef FIO_HAVE_PSHARED_MUTEX
466	if (!o->use_thread) {
467		log_info("fio: this platform does not support process shared"
468			 " mutexes, forcing use of threads. Use the 'thread'"
469			 " option to get rid of this warning.\n");
470		o->use_thread = 1;
471		ret = warnings_fatal;
472	}
473#endif
474
475	if (o->write_iolog_file && o->read_iolog_file) {
476		log_err("fio: read iolog overrides write_iolog\n");
477		free(o->write_iolog_file);
478		o->write_iolog_file = NULL;
479		ret = warnings_fatal;
480	}
481
482	/*
483	 * only really works with 1 file
484	 */
485	if (o->zone_size && o->open_files > 1)
486		o->zone_size = 0;
487
488	/*
489	 * If zone_range isn't specified, backward compatibility dictates it
490	 * should be made equal to zone_size.
491	 */
492	if (o->zone_size && !o->zone_range)
493		o->zone_range = o->zone_size;
494
495	/*
496	 * Reads can do overwrites, we always need to pre-create the file
497	 */
498	if (td_read(td) || td_rw(td))
499		o->overwrite = 1;
500
501	if (!o->min_bs[DDIR_READ])
502		o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
503	if (!o->max_bs[DDIR_READ])
504		o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
505	if (!o->min_bs[DDIR_WRITE])
506		o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
507	if (!o->max_bs[DDIR_WRITE])
508		o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
509	if (!o->min_bs[DDIR_TRIM])
510		o->min_bs[DDIR_TRIM] = o->bs[DDIR_TRIM];
511	if (!o->max_bs[DDIR_TRIM])
512		o->max_bs[DDIR_TRIM] = o->bs[DDIR_TRIM];
513
514
515	o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
516	o->rw_min_bs = min(o->min_bs[DDIR_TRIM], o->rw_min_bs);
517
518	/*
519	 * For random IO, allow blockalign offset other than min_bs.
520	 */
521	if (!o->ba[DDIR_READ] || !td_random(td))
522		o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
523	if (!o->ba[DDIR_WRITE] || !td_random(td))
524		o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
525	if (!o->ba[DDIR_TRIM] || !td_random(td))
526		o->ba[DDIR_TRIM] = o->min_bs[DDIR_TRIM];
527
528	if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
529	    o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE] ||
530	    o->ba[DDIR_TRIM] != o->min_bs[DDIR_TRIM]) &&
531	    !o->norandommap) {
532		log_err("fio: Any use of blockalign= turns off randommap\n");
533		o->norandommap = 1;
534		ret = warnings_fatal;
535	}
536
537	if (!o->file_size_high)
538		o->file_size_high = o->file_size_low;
539
540	if (o->start_delay_high)
541		o->start_delay = get_rand_start_delay(td);
542
543	if (o->norandommap && o->verify != VERIFY_NONE
544	    && !fixed_block_size(o))  {
545		log_err("fio: norandommap given for variable block sizes, "
546			"verify disabled\n");
547		o->verify = VERIFY_NONE;
548		ret = warnings_fatal;
549	}
550	if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
551		log_err("fio: bs_unaligned may not work with raw io\n");
552
553	/*
554	 * thinktime_spin must be less than thinktime
555	 */
556	if (o->thinktime_spin > o->thinktime)
557		o->thinktime_spin = o->thinktime;
558
559	/*
560	 * The low water mark cannot be bigger than the iodepth
561	 */
562	if (o->iodepth_low > o->iodepth || !o->iodepth_low)
563		o->iodepth_low = o->iodepth;
564
565	/*
566	 * If batch number isn't set, default to the same as iodepth
567	 */
568	if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
569		o->iodepth_batch = o->iodepth;
570
571	if (o->nr_files > td->files_index)
572		o->nr_files = td->files_index;
573
574	if (o->open_files > o->nr_files || !o->open_files)
575		o->open_files = o->nr_files;
576
577	if (((o->rate[DDIR_READ] + o->rate[DDIR_WRITE] + o->rate[DDIR_TRIM]) &&
578	    (o->rate_iops[DDIR_READ] + o->rate_iops[DDIR_WRITE] + o->rate_iops[DDIR_TRIM])) ||
579	    ((o->ratemin[DDIR_READ] + o->ratemin[DDIR_WRITE] + o->ratemin[DDIR_TRIM]) &&
580	    (o->rate_iops_min[DDIR_READ] + o->rate_iops_min[DDIR_WRITE] + o->rate_iops_min[DDIR_TRIM]))) {
581		log_err("fio: rate and rate_iops are mutually exclusive\n");
582		ret = 1;
583	}
584	if ((o->rate[DDIR_READ] < o->ratemin[DDIR_READ]) ||
585	    (o->rate[DDIR_WRITE] < o->ratemin[DDIR_WRITE]) ||
586	    (o->rate[DDIR_TRIM] < o->ratemin[DDIR_TRIM]) ||
587	    (o->rate_iops[DDIR_READ] < o->rate_iops_min[DDIR_READ]) ||
588	    (o->rate_iops[DDIR_WRITE] < o->rate_iops_min[DDIR_WRITE]) ||
589	    (o->rate_iops[DDIR_TRIM] < o->rate_iops_min[DDIR_TRIM])) {
590		log_err("fio: minimum rate exceeds rate\n");
591		ret = 1;
592	}
593
594	if (!o->timeout && o->time_based) {
595		log_err("fio: time_based requires a runtime/timeout setting\n");
596		o->time_based = 0;
597		ret = warnings_fatal;
598	}
599
600	if (o->fill_device && !o->size)
601		o->size = -1ULL;
602
603	if (o->verify != VERIFY_NONE) {
604		if (td_write(td) && o->do_verify && o->numjobs > 1) {
605			log_info("Multiple writers may overwrite blocks that "
606				"belong to other jobs. This can cause "
607				"verification failures.\n");
608			ret = warnings_fatal;
609		}
610
611		o->refill_buffers = 1;
612		if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
613		    !o->verify_interval)
614			o->verify_interval = o->min_bs[DDIR_WRITE];
615	}
616
617	if (o->pre_read) {
618		o->invalidate_cache = 0;
619		if (td->io_ops->flags & FIO_PIPEIO) {
620			log_info("fio: cannot pre-read files with an IO engine"
621				 " that isn't seekable. Pre-read disabled.\n");
622			ret = warnings_fatal;
623		}
624	}
625
626	if (!o->unit_base) {
627		if (td->io_ops->flags & FIO_BIT_BASED)
628			o->unit_base = 1;
629		else
630			o->unit_base = 8;
631	}
632
633#ifndef CONFIG_FDATASYNC
634	if (o->fdatasync_blocks) {
635		log_info("fio: this platform does not support fdatasync()"
636			 " falling back to using fsync().  Use the 'fsync'"
637			 " option instead of 'fdatasync' to get rid of"
638			 " this warning\n");
639		o->fsync_blocks = o->fdatasync_blocks;
640		o->fdatasync_blocks = 0;
641		ret = warnings_fatal;
642	}
643#endif
644
645#ifdef WIN32
646	/*
647	 * Windows doesn't support O_DIRECT or O_SYNC with the _open interface,
648	 * so fail if we're passed those flags
649	 */
650	if ((td->io_ops->flags & FIO_SYNCIO) && (td->o.odirect || td->o.sync_io)) {
651		log_err("fio: Windows does not support direct or non-buffered io with"
652				" the synchronous ioengines. Use the 'windowsaio' ioengine"
653				" with 'direct=1' and 'iodepth=1' instead.\n");
654		ret = 1;
655	}
656#endif
657
658	/*
659	 * For fully compressible data, just zero them at init time.
660	 * It's faster than repeatedly filling it.
661	 */
662	if (td->o.compress_percentage == 100) {
663		td->o.zero_buffers = 1;
664		td->o.compress_percentage = 0;
665	}
666
667	/*
668	 * Using a non-uniform random distribution excludes usage of
669	 * a random map
670	 */
671	if (td->o.random_distribution != FIO_RAND_DIST_RANDOM)
672		td->o.norandommap = 1;
673
674	/*
675	 * If size is set but less than the min block size, complain
676	 */
677	if (o->size && o->size < td_min_bs(td)) {
678		log_err("fio: size too small, must be larger than the IO size: %llu\n", (unsigned long long) o->size);
679		ret = 1;
680	}
681
682	/*
683	 * O_ATOMIC implies O_DIRECT
684	 */
685	if (td->o.oatomic)
686		td->o.odirect = 1;
687
688	/*
689	 * If randseed is set, that overrides randrepeat
690	 */
691	if (td->o.rand_seed)
692		td->o.rand_repeatable = 0;
693
694	if ((td->io_ops->flags & FIO_NOEXTEND) && td->o.file_append) {
695		log_err("fio: can't append/extent with IO engine %s\n", td->io_ops->name);
696		ret = 1;
697	}
698
699	return ret;
700}
701
702/*
703 * This function leaks the buffer
704 */
705char *fio_uint_to_kmg(unsigned int val)
706{
707	char *buf = malloc(32);
708	char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
709	char *p = post;
710
711	do {
712		if (val & 1023)
713			break;
714
715		val >>= 10;
716		p++;
717	} while (*p);
718
719	snprintf(buf, 32, "%u%c", val, *p);
720	return buf;
721}
722
723/* External engines are specified by "external:name.o") */
724static const char *get_engine_name(const char *str)
725{
726	char *p = strstr(str, ":");
727
728	if (!p)
729		return str;
730
731	p++;
732	strip_blank_front(&p);
733	strip_blank_end(p);
734	return p;
735}
736
737static int exists_and_not_file(const char *filename)
738{
739	struct stat sb;
740
741	if (lstat(filename, &sb) == -1)
742		return 0;
743
744	/* \\.\ is the device namespace in Windows, where every file
745	 * is a device node */
746	if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
747		return 0;
748
749	return 1;
750}
751
752static void td_fill_rand_seeds_os(struct thread_data *td)
753{
754	os_random_seed(td->rand_seeds[FIO_RAND_BS_OFF], &td->bsrange_state);
755	os_random_seed(td->rand_seeds[FIO_RAND_VER_OFF], &td->verify_state);
756	os_random_seed(td->rand_seeds[FIO_RAND_MIX_OFF], &td->rwmix_state);
757
758	if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
759		os_random_seed(td->rand_seeds[FIO_RAND_FILE_OFF], &td->next_file_state);
760
761	os_random_seed(td->rand_seeds[FIO_RAND_FILE_SIZE_OFF], &td->file_size_state);
762	os_random_seed(td->rand_seeds[FIO_RAND_TRIM_OFF], &td->trim_state);
763	os_random_seed(td->rand_seeds[FIO_RAND_START_DELAY], &td->delay_state);
764
765	if (!td_random(td))
766		return;
767
768	if (td->o.rand_repeatable)
769		td->rand_seeds[FIO_RAND_BLOCK_OFF] = FIO_RANDSEED * td->thread_number;
770
771	os_random_seed(td->rand_seeds[FIO_RAND_BLOCK_OFF], &td->random_state);
772
773	os_random_seed(td->rand_seeds[FIO_RAND_SEQ_RAND_READ_OFF], &td->seq_rand_state[DDIR_READ]);
774	os_random_seed(td->rand_seeds[FIO_RAND_SEQ_RAND_WRITE_OFF], &td->seq_rand_state[DDIR_WRITE]);
775	os_random_seed(td->rand_seeds[FIO_RAND_SEQ_RAND_TRIM_OFF], &td->seq_rand_state[DDIR_TRIM]);
776}
777
778static void td_fill_rand_seeds_internal(struct thread_data *td)
779{
780	init_rand_seed(&td->__bsrange_state, td->rand_seeds[FIO_RAND_BS_OFF]);
781	init_rand_seed(&td->__verify_state, td->rand_seeds[FIO_RAND_VER_OFF]);
782	init_rand_seed(&td->__rwmix_state, td->rand_seeds[FIO_RAND_MIX_OFF]);
783
784	if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
785		init_rand_seed(&td->__next_file_state, td->rand_seeds[FIO_RAND_FILE_OFF]);
786
787	init_rand_seed(&td->__file_size_state, td->rand_seeds[FIO_RAND_FILE_SIZE_OFF]);
788	init_rand_seed(&td->__trim_state, td->rand_seeds[FIO_RAND_TRIM_OFF]);
789	init_rand_seed(&td->__delay_state, td->rand_seeds[FIO_RAND_START_DELAY]);
790
791	if (!td_random(td))
792		return;
793
794	if (td->o.rand_repeatable)
795		td->rand_seeds[FIO_RAND_BLOCK_OFF] = FIO_RANDSEED * td->thread_number;
796
797	init_rand_seed(&td->__random_state, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
798	init_rand_seed(&td->__seq_rand_state[DDIR_READ], td->rand_seeds[FIO_RAND_SEQ_RAND_READ_OFF]);
799	init_rand_seed(&td->__seq_rand_state[DDIR_WRITE], td->rand_seeds[FIO_RAND_SEQ_RAND_WRITE_OFF]);
800	init_rand_seed(&td->__seq_rand_state[DDIR_TRIM], td->rand_seeds[FIO_RAND_SEQ_RAND_TRIM_OFF]);
801}
802
803void td_fill_rand_seeds(struct thread_data *td)
804{
805	if (td->o.allrand_repeatable) {
806		for (int i = 0; i < FIO_RAND_NR_OFFS; i++)
807			td->rand_seeds[i] = FIO_RANDSEED * td->thread_number
808			       	+ i;
809	}
810
811	if (td->o.use_os_rand)
812		td_fill_rand_seeds_os(td);
813	else
814		td_fill_rand_seeds_internal(td);
815
816	init_rand_seed(&td->buf_state, td->rand_seeds[FIO_RAND_BUF_OFF]);
817}
818
819/*
820 * Initializes the ioengine configured for a job, if it has not been done so
821 * already.
822 */
823int ioengine_load(struct thread_data *td)
824{
825	const char *engine;
826
827	/*
828	 * Engine has already been loaded.
829	 */
830	if (td->io_ops)
831		return 0;
832	if (!td->o.ioengine) {
833		log_err("fio: internal fault, no IO engine specified\n");
834		return 1;
835	}
836
837	engine = get_engine_name(td->o.ioengine);
838	td->io_ops = load_ioengine(td, engine);
839	if (!td->io_ops) {
840		log_err("fio: failed to load engine %s\n", engine);
841		return 1;
842	}
843
844	if (td->io_ops->option_struct_size && td->io_ops->options) {
845		/*
846		 * In cases where td->eo is set, clone it for a child thread.
847		 * This requires that the parent thread has the same ioengine,
848		 * but that requirement must be enforced by the code which
849		 * cloned the thread.
850		 */
851		void *origeo = td->eo;
852		/*
853		 * Otherwise use the default thread options.
854		 */
855		if (!origeo && td != &def_thread && def_thread.eo &&
856		    def_thread.io_ops->options == td->io_ops->options)
857			origeo = def_thread.eo;
858
859		options_init(td->io_ops->options);
860		td->eo = malloc(td->io_ops->option_struct_size);
861		/*
862		 * Use the default thread as an option template if this uses the
863		 * same options structure and there are non-default options
864		 * used.
865		 */
866		if (origeo) {
867			memcpy(td->eo, origeo, td->io_ops->option_struct_size);
868			options_mem_dupe(td->eo, td->io_ops->options);
869		} else {
870			memset(td->eo, 0, td->io_ops->option_struct_size);
871			fill_default_options(td->eo, td->io_ops->options);
872		}
873		*(struct thread_data **)td->eo = td;
874	}
875
876	return 0;
877}
878
879static void init_flags(struct thread_data *td)
880{
881	struct thread_options *o = &td->o;
882
883	if (o->verify_backlog)
884		td->flags |= TD_F_VER_BACKLOG;
885	if (o->trim_backlog)
886		td->flags |= TD_F_TRIM_BACKLOG;
887	if (o->read_iolog_file)
888		td->flags |= TD_F_READ_IOLOG;
889	if (o->refill_buffers)
890		td->flags |= TD_F_REFILL_BUFFERS;
891	if (o->scramble_buffers)
892		td->flags |= TD_F_SCRAMBLE_BUFFERS;
893	if (o->verify != VERIFY_NONE)
894		td->flags |= TD_F_VER_NONE;
895}
896
897static int setup_random_seeds(struct thread_data *td)
898{
899	unsigned long seed;
900	unsigned int i;
901
902	if (!td->o.rand_repeatable && !td->o.rand_seed)
903		return init_random_state(td, td->rand_seeds, sizeof(td->rand_seeds));
904
905	if (!td->o.rand_seed)
906		seed = 0x89;
907	else
908		seed = td->o.rand_seed;
909
910	for (i = 0; i < 4; i++)
911		seed *= 0x9e370001UL;
912
913	for (i = 0; i < FIO_RAND_NR_OFFS; i++) {
914		td->rand_seeds[i] = seed;
915		seed *= 0x9e370001UL;
916	}
917
918	td_fill_rand_seeds(td);
919	return 0;
920}
921
922enum {
923	FPRE_NONE = 0,
924	FPRE_JOBNAME,
925	FPRE_JOBNUM,
926	FPRE_FILENUM
927};
928
929static struct fpre_keyword {
930	const char *keyword;
931	size_t strlen;
932	int key;
933} fpre_keywords[] = {
934	{ .keyword = "$jobname",	.key = FPRE_JOBNAME, },
935	{ .keyword = "$jobnum",		.key = FPRE_JOBNUM, },
936	{ .keyword = "$filenum",	.key = FPRE_FILENUM, },
937	{ .keyword = NULL, },
938	};
939
940static char *make_filename(char *buf, struct thread_options *o,
941			   const char *jobname, int jobnum, int filenum)
942{
943	struct fpre_keyword *f;
944	char copy[PATH_MAX];
945	size_t dst_left = PATH_MAX - 1;
946
947	if (!o->filename_format || !strlen(o->filename_format)) {
948		sprintf(buf, "%s.%d.%d", jobname, jobnum, filenum);
949		return NULL;
950	}
951
952	for (f = &fpre_keywords[0]; f->keyword; f++)
953		f->strlen = strlen(f->keyword);
954
955	strcpy(buf, o->filename_format);
956	memset(copy, 0, sizeof(copy));
957	for (f = &fpre_keywords[0]; f->keyword; f++) {
958		do {
959			size_t pre_len, post_start = 0;
960			char *str, *dst = copy;
961
962			str = strcasestr(buf, f->keyword);
963			if (!str)
964				break;
965
966			pre_len = str - buf;
967			if (strlen(str) != f->strlen)
968				post_start = pre_len + f->strlen;
969
970			if (pre_len) {
971				strncpy(dst, buf, pre_len);
972				dst += pre_len;
973				dst_left -= pre_len;
974			}
975
976			switch (f->key) {
977			case FPRE_JOBNAME: {
978				int ret;
979
980				ret = snprintf(dst, dst_left, "%s", jobname);
981				if (ret < 0)
982					break;
983				dst += ret;
984				dst_left -= ret;
985				break;
986				}
987			case FPRE_JOBNUM: {
988				int ret;
989
990				ret = snprintf(dst, dst_left, "%d", jobnum);
991				if (ret < 0)
992					break;
993				dst += ret;
994				dst_left -= ret;
995				break;
996				}
997			case FPRE_FILENUM: {
998				int ret;
999
1000				ret = snprintf(dst, dst_left, "%d", filenum);
1001				if (ret < 0)
1002					break;
1003				dst += ret;
1004				dst_left -= ret;
1005				break;
1006				}
1007			default:
1008				assert(0);
1009				break;
1010			}
1011
1012			if (post_start)
1013				strncpy(dst, buf + post_start, dst_left);
1014
1015			strcpy(buf, copy);
1016		} while (1);
1017	}
1018
1019	return buf;
1020}
1021
1022int parse_dryrun(void)
1023{
1024	return dump_cmdline || parse_only;
1025}
1026
1027/*
1028 * Adds a job to the list of things todo. Sanitizes the various options
1029 * to make sure we don't have conflicts, and initializes various
1030 * members of td.
1031 */
1032static int add_job(struct thread_data *td, const char *jobname, int job_add_num,
1033		   int recursed, int client_type)
1034{
1035	unsigned int i;
1036	char fname[PATH_MAX];
1037	int numjobs, file_alloced;
1038	struct thread_options *o = &td->o;
1039
1040	/*
1041	 * the def_thread is just for options, it's not a real job
1042	 */
1043	if (td == &def_thread)
1044		return 0;
1045
1046	init_flags(td);
1047
1048	/*
1049	 * if we are just dumping the output command line, don't add the job
1050	 */
1051	if (parse_dryrun()) {
1052		put_job(td);
1053		return 0;
1054	}
1055
1056	td->client_type = client_type;
1057
1058	if (profile_td_init(td))
1059		goto err;
1060
1061	if (ioengine_load(td))
1062		goto err;
1063
1064	if (o->odirect)
1065		td->io_ops->flags |= FIO_RAWIO;
1066
1067	file_alloced = 0;
1068	if (!o->filename && !td->files_index && !o->read_iolog_file) {
1069		file_alloced = 1;
1070
1071		if (o->nr_files == 1 && exists_and_not_file(jobname))
1072			add_file(td, jobname, job_add_num, 0);
1073		else {
1074			for (i = 0; i < o->nr_files; i++)
1075				add_file(td, make_filename(fname, o, jobname, job_add_num, i), job_add_num, 0);
1076		}
1077	}
1078
1079	if (fixup_options(td))
1080		goto err;
1081
1082	flow_init_job(td);
1083
1084	/*
1085	 * IO engines only need this for option callbacks, and the address may
1086	 * change in subprocesses.
1087	 */
1088	if (td->eo)
1089		*(struct thread_data **)td->eo = NULL;
1090
1091	if (td->io_ops->flags & FIO_DISKLESSIO) {
1092		struct fio_file *f;
1093
1094		for_each_file(td, f, i)
1095			f->real_file_size = -1ULL;
1096	}
1097
1098	td->mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1099
1100	td->ts.clat_percentiles = o->clat_percentiles;
1101	td->ts.percentile_precision = o->percentile_precision;
1102	memcpy(td->ts.percentile_list, o->percentile_list, sizeof(o->percentile_list));
1103
1104	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1105		td->ts.clat_stat[i].min_val = ULONG_MAX;
1106		td->ts.slat_stat[i].min_val = ULONG_MAX;
1107		td->ts.lat_stat[i].min_val = ULONG_MAX;
1108		td->ts.bw_stat[i].min_val = ULONG_MAX;
1109	}
1110	td->ddir_seq_nr = o->ddir_seq_nr;
1111
1112	if ((o->stonewall || o->new_group) && prev_group_jobs) {
1113		prev_group_jobs = 0;
1114		groupid++;
1115	}
1116
1117	td->groupid = groupid;
1118	prev_group_jobs++;
1119
1120	if (setup_random_seeds(td)) {
1121		td_verror(td, errno, "init_random_state");
1122		goto err;
1123	}
1124
1125	if (setup_rate(td))
1126		goto err;
1127
1128	if (o->lat_log_file || write_lat_log) {
1129		setup_log(&td->lat_log, o->log_avg_msec, IO_LOG_TYPE_LAT);
1130		setup_log(&td->slat_log, o->log_avg_msec, IO_LOG_TYPE_SLAT);
1131		setup_log(&td->clat_log, o->log_avg_msec, IO_LOG_TYPE_CLAT);
1132	}
1133	if (o->bw_log_file || write_bw_log)
1134		setup_log(&td->bw_log, o->log_avg_msec, IO_LOG_TYPE_BW);
1135	if (o->iops_log_file)
1136		setup_log(&td->iops_log, o->log_avg_msec, IO_LOG_TYPE_IOPS);
1137
1138	if (!o->name)
1139		o->name = strdup(jobname);
1140
1141	if (output_format == FIO_OUTPUT_NORMAL) {
1142		if (!job_add_num) {
1143			if (is_backend && !recursed)
1144				fio_server_send_add_job(td);
1145
1146			if (!(td->io_ops->flags & FIO_NOIO)) {
1147				char *c1, *c2, *c3, *c4;
1148				char *c5 = NULL, *c6 = NULL;
1149
1150				c1 = fio_uint_to_kmg(o->min_bs[DDIR_READ]);
1151				c2 = fio_uint_to_kmg(o->max_bs[DDIR_READ]);
1152				c3 = fio_uint_to_kmg(o->min_bs[DDIR_WRITE]);
1153				c4 = fio_uint_to_kmg(o->max_bs[DDIR_WRITE]);
1154
1155				if (!o->bs_is_seq_rand) {
1156					c5 = fio_uint_to_kmg(o->min_bs[DDIR_TRIM]);
1157					c6 = fio_uint_to_kmg(o->max_bs[DDIR_TRIM]);
1158				}
1159
1160				log_info("%s: (g=%d): rw=%s, ", td->o.name,
1161							td->groupid,
1162							ddir_str(o->td_ddir));
1163
1164				if (o->bs_is_seq_rand)
1165					log_info("bs(seq/rand)=%s-%s/%s-%s, ",
1166							c1, c2, c3, c4);
1167				else
1168					log_info("bs=%s-%s/%s-%s/%s-%s, ",
1169							c1, c2, c3, c4, c5, c6);
1170
1171				log_info("ioengine=%s, iodepth=%u\n",
1172						td->io_ops->name, o->iodepth);
1173
1174				free(c1);
1175				free(c2);
1176				free(c3);
1177				free(c4);
1178				free(c5);
1179				free(c6);
1180			}
1181		} else if (job_add_num == 1)
1182			log_info("...\n");
1183	}
1184
1185	/*
1186	 * recurse add identical jobs, clear numjobs and stonewall options
1187	 * as they don't apply to sub-jobs
1188	 */
1189	numjobs = o->numjobs;
1190	while (--numjobs) {
1191		struct thread_data *td_new = get_new_job(0, td, 1);
1192
1193		if (!td_new)
1194			goto err;
1195
1196		td_new->o.numjobs = 1;
1197		td_new->o.stonewall = 0;
1198		td_new->o.new_group = 0;
1199
1200		if (file_alloced) {
1201			if (td_new->files) {
1202				struct fio_file *f;
1203				for_each_file(td_new, f, i) {
1204					if (f->file_name)
1205						sfree(f->file_name);
1206					sfree(f);
1207				}
1208				free(td_new->files);
1209				td_new->files = NULL;
1210			}
1211			td_new->files_index = 0;
1212			td_new->files_size = 0;
1213			if (td_new->o.filename) {
1214				free(td_new->o.filename);
1215				td_new->o.filename = NULL;
1216			}
1217		}
1218
1219		if (add_job(td_new, jobname, numjobs, 1, client_type))
1220			goto err;
1221	}
1222
1223	return 0;
1224err:
1225	put_job(td);
1226	return -1;
1227}
1228
1229/*
1230 * Parse as if 'o' was a command line
1231 */
1232void add_job_opts(const char **o, int client_type)
1233{
1234	struct thread_data *td, *td_parent;
1235	int i, in_global = 1;
1236	char jobname[32];
1237
1238	i = 0;
1239	td_parent = td = NULL;
1240	while (o[i]) {
1241		if (!strncmp(o[i], "name", 4)) {
1242			in_global = 0;
1243			if (td)
1244				add_job(td, jobname, 0, 0, client_type);
1245			td = NULL;
1246			sprintf(jobname, "%s", o[i] + 5);
1247		}
1248		if (in_global && !td_parent)
1249			td_parent = get_new_job(1, &def_thread, 0);
1250		else if (!in_global && !td) {
1251			if (!td_parent)
1252				td_parent = &def_thread;
1253			td = get_new_job(0, td_parent, 0);
1254		}
1255		if (in_global)
1256			fio_options_parse(td_parent, (char **) &o[i], 1, 0);
1257		else
1258			fio_options_parse(td, (char **) &o[i], 1, 0);
1259		i++;
1260	}
1261
1262	if (td)
1263		add_job(td, jobname, 0, 0, client_type);
1264}
1265
1266static int skip_this_section(const char *name)
1267{
1268	int i;
1269
1270	if (!nr_job_sections)
1271		return 0;
1272	if (!strncmp(name, "global", 6))
1273		return 0;
1274
1275	for (i = 0; i < nr_job_sections; i++)
1276		if (!strcmp(job_sections[i], name))
1277			return 0;
1278
1279	return 1;
1280}
1281
1282static int is_empty_or_comment(char *line)
1283{
1284	unsigned int i;
1285
1286	for (i = 0; i < strlen(line); i++) {
1287		if (line[i] == ';')
1288			return 1;
1289		if (line[i] == '#')
1290			return 1;
1291		if (!isspace((int) line[i]) && !iscntrl((int) line[i]))
1292			return 0;
1293	}
1294
1295	return 1;
1296}
1297
1298/*
1299 * This is our [ini] type file parser.
1300 */
1301int parse_jobs_ini(char *file, int is_buf, int stonewall_flag, int type)
1302{
1303	unsigned int global;
1304	struct thread_data *td;
1305	char *string, *name;
1306	FILE *f;
1307	char *p;
1308	int ret = 0, stonewall;
1309	int first_sect = 1;
1310	int skip_fgets = 0;
1311	int inside_skip = 0;
1312	char **opts;
1313	int i, alloc_opts, num_opts;
1314
1315	if (is_buf)
1316		f = NULL;
1317	else {
1318		if (!strcmp(file, "-"))
1319			f = stdin;
1320		else
1321			f = fopen(file, "r");
1322
1323		if (!f) {
1324			perror("fopen job file");
1325			return 1;
1326		}
1327	}
1328
1329	string = malloc(4096);
1330
1331	/*
1332	 * it's really 256 + small bit, 280 should suffice
1333	 */
1334	name = malloc(280);
1335	memset(name, 0, 280);
1336
1337	alloc_opts = 8;
1338	opts = malloc(sizeof(char *) * alloc_opts);
1339	num_opts = 0;
1340
1341	stonewall = stonewall_flag;
1342	do {
1343		/*
1344		 * if skip_fgets is set, we already have loaded a line we
1345		 * haven't handled.
1346		 */
1347		if (!skip_fgets) {
1348			if (is_buf)
1349				p = strsep(&file, "\n");
1350			else
1351				p = fgets(string, 4096, f);
1352			if (!p)
1353				break;
1354		}
1355
1356		skip_fgets = 0;
1357		strip_blank_front(&p);
1358		strip_blank_end(p);
1359
1360		if (is_empty_or_comment(p))
1361			continue;
1362		if (sscanf(p, "[%255[^\n]]", name) != 1) {
1363			if (inside_skip)
1364				continue;
1365			log_err("fio: option <%s> outside of [] job section\n",
1366									p);
1367			break;
1368		}
1369
1370		name[strlen(name) - 1] = '\0';
1371
1372		if (skip_this_section(name)) {
1373			inside_skip = 1;
1374			continue;
1375		} else
1376			inside_skip = 0;
1377
1378		global = !strncmp(name, "global", 6);
1379
1380		if (dump_cmdline) {
1381			if (first_sect)
1382				log_info("fio ");
1383			if (!global)
1384				log_info("--name=%s ", name);
1385			first_sect = 0;
1386		}
1387
1388		td = get_new_job(global, &def_thread, 0);
1389		if (!td) {
1390			ret = 1;
1391			break;
1392		}
1393
1394		/*
1395		 * Separate multiple job files by a stonewall
1396		 */
1397		if (!global && stonewall) {
1398			td->o.stonewall = stonewall;
1399			stonewall = 0;
1400		}
1401
1402		num_opts = 0;
1403		memset(opts, 0, alloc_opts * sizeof(char *));
1404
1405		while (1) {
1406			if (is_buf)
1407				p = strsep(&file, "\n");
1408			else
1409				p = fgets(string, 4096, f);
1410			if (!p)
1411				break;
1412
1413			if (is_empty_or_comment(p))
1414				continue;
1415
1416			strip_blank_front(&p);
1417
1418			/*
1419			 * new section, break out and make sure we don't
1420			 * fgets() a new line at the top.
1421			 */
1422			if (p[0] == '[') {
1423				skip_fgets = 1;
1424				break;
1425			}
1426
1427			strip_blank_end(p);
1428
1429			if (num_opts == alloc_opts) {
1430				alloc_opts <<= 1;
1431				opts = realloc(opts,
1432						alloc_opts * sizeof(char *));
1433			}
1434
1435			opts[num_opts] = strdup(p);
1436			num_opts++;
1437		}
1438
1439		ret = fio_options_parse(td, opts, num_opts, dump_cmdline);
1440		if (!ret)
1441			ret = add_job(td, name, 0, 0, type);
1442		else {
1443			log_err("fio: job %s dropped\n", name);
1444			put_job(td);
1445		}
1446
1447		for (i = 0; i < num_opts; i++)
1448			free(opts[i]);
1449		num_opts = 0;
1450	} while (!ret);
1451
1452	if (dump_cmdline)
1453		log_info("\n");
1454
1455	i = 0;
1456	while (i < nr_job_sections) {
1457		free(job_sections[i]);
1458		i++;
1459	}
1460
1461	free(string);
1462	free(name);
1463	free(opts);
1464	if (!is_buf && f != stdin)
1465		fclose(f);
1466	return ret;
1467}
1468
1469static int fill_def_thread(void)
1470{
1471	memset(&def_thread, 0, sizeof(def_thread));
1472
1473	fio_getaffinity(getpid(), &def_thread.o.cpumask);
1474	def_thread.o.error_dump = 1;
1475
1476	/*
1477	 * fill default options
1478	 */
1479	fio_fill_default_options(&def_thread);
1480	return 0;
1481}
1482
1483static void usage(const char *name)
1484{
1485	printf("%s\n", fio_version_string);
1486	printf("%s [options] [job options] <job file(s)>\n", name);
1487	printf("  --debug=options\tEnable debug logging. May be one/more of:\n"
1488		"\t\t\tprocess,file,io,mem,blktrace,verify,random,parse,\n"
1489		"\t\t\tdiskutil,job,mutex,profile,time,net,rate\n");
1490	printf("  --parse-only\t\tParse options only, don't start any IO\n");
1491	printf("  --output\t\tWrite output to file\n");
1492	printf("  --runtime\t\tRuntime in seconds\n");
1493	printf("  --latency-log\t\tGenerate per-job latency logs\n");
1494	printf("  --bandwidth-log\tGenerate per-job bandwidth logs\n");
1495	printf("  --minimal\t\tMinimal (terse) output\n");
1496	printf("  --output-format=x\tOutput format (terse,json,normal)\n");
1497	printf("  --terse-version=x\tSet terse version output format to 'x'\n");
1498	printf("  --version\t\tPrint version info and exit\n");
1499	printf("  --help\t\tPrint this page\n");
1500	printf("  --cpuclock-test\tPerform test/validation of CPU clock\n");
1501	printf("  --crctest\t\tTest speed of checksum functions\n");
1502	printf("  --cmdhelp=cmd\t\tPrint command help, \"all\" for all of"
1503		" them\n");
1504	printf("  --enghelp=engine\tPrint ioengine help, or list"
1505		" available ioengines\n");
1506	printf("  --enghelp=engine,cmd\tPrint help for an ioengine"
1507		" cmd\n");
1508	printf("  --showcmd\t\tTurn a job file into command line options\n");
1509	printf("  --eta=when\t\tWhen ETA estimate should be printed\n");
1510	printf("            \t\tMay be \"always\", \"never\" or \"auto\"\n");
1511	printf("  --eta-newline=time\tForce a new line for every 'time'");
1512	printf(" period passed\n");
1513	printf("  --status-interval=t\tForce full status dump every");
1514	printf(" 't' period passed\n");
1515	printf("  --readonly\t\tTurn on safety read-only checks, preventing"
1516		" writes\n");
1517	printf("  --section=name\tOnly run specified section in job file\n");
1518	printf("  --alloc-size=kb\tSet smalloc pool to this size in kb"
1519		" (def 1024)\n");
1520	printf("  --warnings-fatal\tFio parser warnings are fatal\n");
1521	printf("  --max-jobs=nr\t\tMaximum number of threads/processes to support\n");
1522	printf("  --server=args\t\tStart a backend fio server\n");
1523	printf("  --daemonize=pidfile\tBackground fio server, write pid to file\n");
1524	printf("  --client=hostname\tTalk to remote backend fio server at hostname\n");
1525	printf("  --idle-prof=option\tReport cpu idleness on a system or percpu basis\n"
1526		"\t\t\t(option=system,percpu) or run unit work\n"
1527		"\t\t\tcalibration only (option=calibrate)\n");
1528	printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
1529	printf("\n                   Jens Axboe <jaxboe@fusionio.com>");
1530	printf("\n                   Jens Axboe <axboe@fb.com>\n");
1531}
1532
1533#ifdef FIO_INC_DEBUG
1534struct debug_level debug_levels[] = {
1535	{ .name = "process",
1536	  .help = "Process creation/exit logging",
1537	  .shift = FD_PROCESS,
1538	},
1539	{ .name = "file",
1540	  .help = "File related action logging",
1541	  .shift = FD_FILE,
1542	},
1543	{ .name = "io",
1544	  .help = "IO and IO engine action logging (offsets, queue, completions, etc)",
1545	  .shift = FD_IO,
1546	},
1547	{ .name = "mem",
1548	  .help = "Memory allocation/freeing logging",
1549	  .shift = FD_MEM,
1550	},
1551	{ .name = "blktrace",
1552	  .help = "blktrace action logging",
1553	  .shift = FD_BLKTRACE,
1554	},
1555	{ .name = "verify",
1556	  .help = "IO verification action logging",
1557	  .shift = FD_VERIFY,
1558	},
1559	{ .name = "random",
1560	  .help = "Random generation logging",
1561	  .shift = FD_RANDOM,
1562	},
1563	{ .name = "parse",
1564	  .help = "Parser logging",
1565	  .shift = FD_PARSE,
1566	},
1567	{ .name = "diskutil",
1568	  .help = "Disk utility logging actions",
1569	  .shift = FD_DISKUTIL,
1570	},
1571	{ .name = "job",
1572	  .help = "Logging related to creating/destroying jobs",
1573	  .shift = FD_JOB,
1574	},
1575	{ .name = "mutex",
1576	  .help = "Mutex logging",
1577	  .shift = FD_MUTEX
1578	},
1579	{ .name	= "profile",
1580	  .help = "Logging related to profiles",
1581	  .shift = FD_PROFILE,
1582	},
1583	{ .name = "time",
1584	  .help = "Logging related to time keeping functions",
1585	  .shift = FD_TIME,
1586	},
1587	{ .name = "net",
1588	  .help = "Network logging",
1589	  .shift = FD_NET,
1590	},
1591	{ .name = "rate",
1592	  .help = "Rate logging",
1593	  .shift = FD_RATE,
1594	},
1595	{ .name = NULL, },
1596};
1597
1598static int set_debug(const char *string)
1599{
1600	struct debug_level *dl;
1601	char *p = (char *) string;
1602	char *opt;
1603	int i;
1604
1605	if (!strcmp(string, "?") || !strcmp(string, "help")) {
1606		log_info("fio: dumping debug options:");
1607		for (i = 0; debug_levels[i].name; i++) {
1608			dl = &debug_levels[i];
1609			log_info("%s,", dl->name);
1610		}
1611		log_info("all\n");
1612		return 1;
1613	}
1614
1615	while ((opt = strsep(&p, ",")) != NULL) {
1616		int found = 0;
1617
1618		if (!strncmp(opt, "all", 3)) {
1619			log_info("fio: set all debug options\n");
1620			fio_debug = ~0UL;
1621			continue;
1622		}
1623
1624		for (i = 0; debug_levels[i].name; i++) {
1625			dl = &debug_levels[i];
1626			found = !strncmp(opt, dl->name, strlen(dl->name));
1627			if (!found)
1628				continue;
1629
1630			if (dl->shift == FD_JOB) {
1631				opt = strchr(opt, ':');
1632				if (!opt) {
1633					log_err("fio: missing job number\n");
1634					break;
1635				}
1636				opt++;
1637				fio_debug_jobno = atoi(opt);
1638				log_info("fio: set debug jobno %d\n",
1639							fio_debug_jobno);
1640			} else {
1641				log_info("fio: set debug option %s\n", opt);
1642				fio_debug |= (1UL << dl->shift);
1643			}
1644			break;
1645		}
1646
1647		if (!found)
1648			log_err("fio: debug mask %s not found\n", opt);
1649	}
1650	return 0;
1651}
1652#else
1653static int set_debug(const char *string)
1654{
1655	log_err("fio: debug tracing not included in build\n");
1656	return 1;
1657}
1658#endif
1659
1660static void fio_options_fill_optstring(void)
1661{
1662	char *ostr = cmd_optstr;
1663	int i, c;
1664
1665	c = i = 0;
1666	while (l_opts[i].name) {
1667		ostr[c++] = l_opts[i].val;
1668		if (l_opts[i].has_arg == required_argument)
1669			ostr[c++] = ':';
1670		else if (l_opts[i].has_arg == optional_argument) {
1671			ostr[c++] = ':';
1672			ostr[c++] = ':';
1673		}
1674		i++;
1675	}
1676	ostr[c] = '\0';
1677}
1678
1679static int client_flag_set(char c)
1680{
1681	int i;
1682
1683	i = 0;
1684	while (l_opts[i].name) {
1685		int val = l_opts[i].val;
1686
1687		if (c == (val & 0xff))
1688			return (val & FIO_CLIENT_FLAG);
1689
1690		i++;
1691	}
1692
1693	return 0;
1694}
1695
1696static void parse_cmd_client(void *client, char *opt)
1697{
1698	fio_client_add_cmd_option(client, opt);
1699}
1700
1701int parse_cmd_line(int argc, char *argv[], int client_type)
1702{
1703	struct thread_data *td = NULL;
1704	int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1705	char *ostr = cmd_optstr;
1706	void *pid_file = NULL;
1707	void *cur_client = NULL;
1708	int backend = 0;
1709
1710	/*
1711	 * Reset optind handling, since we may call this multiple times
1712	 * for the backend.
1713	 */
1714	optind = 1;
1715
1716	while ((c = getopt_long_only(argc, argv, ostr, l_opts, &lidx)) != -1) {
1717		did_arg = 1;
1718
1719		if ((c & FIO_CLIENT_FLAG) || client_flag_set(c)) {
1720			parse_cmd_client(cur_client, argv[optind - 1]);
1721			c &= ~FIO_CLIENT_FLAG;
1722		}
1723
1724		switch (c) {
1725		case 'a':
1726			smalloc_pool_size = atoi(optarg);
1727			break;
1728		case 't':
1729			if (check_str_time(optarg, &def_timeout, 1)) {
1730				log_err("fio: failed parsing time %s\n", optarg);
1731				do_exit++;
1732				exit_val = 1;
1733			}
1734			break;
1735		case 'l':
1736			write_lat_log = 1;
1737			break;
1738		case 'b':
1739			write_bw_log = 1;
1740			break;
1741		case 'o':
1742			if (f_out)
1743				fclose(f_out);
1744
1745			f_out = fopen(optarg, "w+");
1746			if (!f_out) {
1747				perror("fopen output");
1748				exit(1);
1749			}
1750			f_err = f_out;
1751			break;
1752		case 'm':
1753			output_format = FIO_OUTPUT_TERSE;
1754			break;
1755		case 'F':
1756			if (!optarg) {
1757				log_err("fio: missing --output-format argument\n");
1758				exit_val = 1;
1759				do_exit++;
1760				break;
1761			}
1762			if (!strcmp(optarg, "minimal") ||
1763			    !strcmp(optarg, "terse") ||
1764			    !strcmp(optarg, "csv"))
1765				output_format = FIO_OUTPUT_TERSE;
1766			else if (!strcmp(optarg, "json"))
1767				output_format = FIO_OUTPUT_JSON;
1768			else
1769				output_format = FIO_OUTPUT_NORMAL;
1770			break;
1771		case 'f':
1772			append_terse_output = 1;
1773			break;
1774		case 'h':
1775			if (!cur_client) {
1776				usage(argv[0]);
1777				do_exit++;
1778			}
1779			break;
1780		case 'c':
1781			if (!cur_client) {
1782				fio_show_option_help(optarg);
1783				do_exit++;
1784			}
1785			break;
1786		case 'i':
1787			if (!cur_client) {
1788				fio_show_ioengine_help(optarg);
1789				do_exit++;
1790			}
1791			break;
1792		case 's':
1793			dump_cmdline = 1;
1794			break;
1795		case 'r':
1796			read_only = 1;
1797			break;
1798		case 'v':
1799			if (!cur_client) {
1800				log_info("%s\n", fio_version_string);
1801				do_exit++;
1802			}
1803			break;
1804		case 'V':
1805			terse_version = atoi(optarg);
1806			if (!(terse_version == 2 || terse_version == 3 ||
1807			     terse_version == 4)) {
1808				log_err("fio: bad terse version format\n");
1809				exit_val = 1;
1810				do_exit++;
1811			}
1812			break;
1813		case 'e':
1814			if (!strcmp("always", optarg))
1815				eta_print = FIO_ETA_ALWAYS;
1816			else if (!strcmp("never", optarg))
1817				eta_print = FIO_ETA_NEVER;
1818			break;
1819		case 'E': {
1820			long long t = 0;
1821
1822			if (str_to_decimal(optarg, &t, 0, NULL, 1)) {
1823				log_err("fio: failed parsing eta time %s\n", optarg);
1824				exit_val = 1;
1825				do_exit++;
1826			}
1827			eta_new_line = t;
1828			break;
1829			}
1830		case 'd':
1831			if (set_debug(optarg))
1832				do_exit++;
1833			break;
1834		case 'P':
1835			parse_only = 1;
1836			break;
1837		case 'x': {
1838			size_t new_size;
1839
1840			if (!strcmp(optarg, "global")) {
1841				log_err("fio: can't use global as only "
1842					"section\n");
1843				do_exit++;
1844				exit_val = 1;
1845				break;
1846			}
1847			new_size = (nr_job_sections + 1) * sizeof(char *);
1848			job_sections = realloc(job_sections, new_size);
1849			job_sections[nr_job_sections] = strdup(optarg);
1850			nr_job_sections++;
1851			break;
1852			}
1853		case 'p':
1854			if (exec_profile)
1855				free(exec_profile);
1856			exec_profile = strdup(optarg);
1857			break;
1858		case FIO_GETOPT_JOB: {
1859			const char *opt = l_opts[lidx].name;
1860			char *val = optarg;
1861
1862			if (!strncmp(opt, "name", 4) && td) {
1863				ret = add_job(td, td->o.name ?: "fio", 0, 0, client_type);
1864				if (ret)
1865					goto out_free;
1866				td = NULL;
1867			}
1868			if (!td) {
1869				int is_section = !strncmp(opt, "name", 4);
1870				int global = 0;
1871
1872				if (!is_section || !strncmp(val, "global", 6))
1873					global = 1;
1874
1875				if (is_section && skip_this_section(val))
1876					continue;
1877
1878				td = get_new_job(global, &def_thread, 1);
1879				if (!td || ioengine_load(td))
1880					goto out_free;
1881				fio_options_set_ioengine_opts(l_opts, td);
1882			}
1883
1884			if ((!val || !strlen(val)) &&
1885			    l_opts[lidx].has_arg == required_argument) {
1886				log_err("fio: option %s requires an argument\n", opt);
1887				ret = 1;
1888			} else
1889				ret = fio_cmd_option_parse(td, opt, val);
1890
1891			if (ret) {
1892				if (td) {
1893					put_job(td);
1894					td = NULL;
1895				}
1896				do_exit++;
1897			}
1898
1899			if (!ret && !strcmp(opt, "ioengine")) {
1900				free_ioengine(td);
1901				if (ioengine_load(td))
1902					goto out_free;
1903				fio_options_set_ioengine_opts(l_opts, td);
1904			}
1905			break;
1906		}
1907		case FIO_GETOPT_IOENGINE: {
1908			const char *opt = l_opts[lidx].name;
1909			char *val = optarg;
1910
1911			if (!td)
1912				break;
1913
1914			ret = fio_cmd_ioengine_option_parse(td, opt, val);
1915			break;
1916		}
1917		case 'w':
1918			warnings_fatal = 1;
1919			break;
1920		case 'j':
1921			max_jobs = atoi(optarg);
1922			if (!max_jobs || max_jobs > REAL_MAX_JOBS) {
1923				log_err("fio: invalid max jobs: %d\n", max_jobs);
1924				do_exit++;
1925				exit_val = 1;
1926			}
1927			break;
1928		case 'S':
1929			if (nr_clients) {
1930				log_err("fio: can't be both client and server\n");
1931				do_exit++;
1932				exit_val = 1;
1933				break;
1934			}
1935			if (optarg)
1936				fio_server_set_arg(optarg);
1937			is_backend = 1;
1938			backend = 1;
1939			break;
1940		case 'D':
1941			if (pid_file)
1942				free(pid_file);
1943			pid_file = strdup(optarg);
1944			break;
1945		case 'I':
1946			if ((ret = fio_idle_prof_parse_opt(optarg))) {
1947				/* exit on error and calibration only */
1948				do_exit++;
1949				if (ret == -1)
1950					exit_val = 1;
1951			}
1952			break;
1953		case 'C':
1954			if (is_backend) {
1955				log_err("fio: can't be both client and server\n");
1956				do_exit++;
1957				exit_val = 1;
1958				break;
1959			}
1960			if (fio_client_add(&fio_client_ops, optarg, &cur_client)) {
1961				log_err("fio: failed adding client %s\n", optarg);
1962				do_exit++;
1963				exit_val = 1;
1964				break;
1965			}
1966			/*
1967			 * If the next argument exists and isn't an option,
1968			 * assume it's a job file for this client only.
1969			 */
1970			while (optind < argc) {
1971				if (!strncmp(argv[optind], "--", 2) ||
1972				    !strncmp(argv[optind], "-", 1))
1973					break;
1974
1975				fio_client_add_ini_file(cur_client, argv[optind]);
1976				optind++;
1977			}
1978			break;
1979		case 'T':
1980			do_exit++;
1981			exit_val = fio_monotonic_clocktest();
1982			break;
1983		case 'G':
1984			do_exit++;
1985			exit_val = fio_crctest(optarg);
1986			break;
1987		case 'L': {
1988			long long val;
1989
1990			if (check_str_time(optarg, &val, 0)) {
1991				log_err("fio: failed parsing time %s\n", optarg);
1992				do_exit++;
1993				exit_val = 1;
1994				break;
1995			}
1996			status_interval = val * 1000;
1997			break;
1998			}
1999		case '?':
2000			log_err("%s: unrecognized option '%s'\n", argv[0],
2001							argv[optind - 1]);
2002		default:
2003			do_exit++;
2004			exit_val = 1;
2005			break;
2006		}
2007		if (do_exit)
2008			break;
2009	}
2010
2011	if (do_exit && !(is_backend || nr_clients))
2012		exit(exit_val);
2013
2014	if (nr_clients && fio_clients_connect())
2015		exit(1);
2016
2017	if (is_backend && backend)
2018		return fio_start_server(pid_file);
2019	else if (pid_file)
2020		free(pid_file);
2021
2022	if (td) {
2023		if (!ret)
2024			ret = add_job(td, td->o.name ?: "fio", 0, 0, client_type);
2025	}
2026
2027	while (!ret && optind < argc) {
2028		ini_idx++;
2029		ini_file = realloc(ini_file, ini_idx * sizeof(char *));
2030		ini_file[ini_idx - 1] = strdup(argv[optind]);
2031		optind++;
2032	}
2033
2034out_free:
2035	if (pid_file)
2036		free(pid_file);
2037
2038	return ini_idx;
2039}
2040
2041int fio_init_options(void)
2042{
2043	f_out = stdout;
2044	f_err = stderr;
2045
2046	fio_options_fill_optstring();
2047	fio_options_dup_and_init(l_opts);
2048
2049	atexit(free_shm);
2050
2051	if (fill_def_thread())
2052		return 1;
2053
2054	return 0;
2055}
2056
2057extern int fio_check_options(struct thread_options *);
2058
2059int parse_options(int argc, char *argv[])
2060{
2061	const int type = FIO_CLIENT_TYPE_CLI;
2062	int job_files, i;
2063
2064	if (fio_init_options())
2065		return 1;
2066	if (fio_test_cconv(&def_thread.o))
2067		log_err("fio: failed internal cconv test\n");
2068
2069	job_files = parse_cmd_line(argc, argv, type);
2070
2071	if (job_files > 0) {
2072		for (i = 0; i < job_files; i++) {
2073			if (i && fill_def_thread())
2074				return 1;
2075			if (nr_clients) {
2076				if (fio_clients_send_ini(ini_file[i]))
2077					return 1;
2078				free(ini_file[i]);
2079			} else if (!is_backend) {
2080				if (parse_jobs_ini(ini_file[i], 0, i, type))
2081					return 1;
2082				free(ini_file[i]);
2083			}
2084		}
2085	} else if (nr_clients) {
2086		if (fill_def_thread())
2087			return 1;
2088		if (fio_clients_send_ini(NULL))
2089			return 1;
2090	}
2091
2092	free(ini_file);
2093	fio_options_free(&def_thread);
2094	filesetup_mem_free();
2095
2096	if (!thread_number) {
2097		if (parse_dryrun())
2098			return 0;
2099		if (exec_profile)
2100			return 0;
2101		if (is_backend || nr_clients)
2102			return 0;
2103		if (did_arg)
2104			return 0;
2105
2106		log_err("No jobs(s) defined\n\n");
2107
2108		if (!did_arg) {
2109			usage(argv[0]);
2110			return 1;
2111		}
2112
2113		return 0;
2114	}
2115
2116	if (def_thread.o.gtod_offload) {
2117		fio_gtod_init();
2118		fio_gtod_offload = 1;
2119		fio_gtod_cpu = def_thread.o.gtod_cpu;
2120	}
2121
2122	if (output_format == FIO_OUTPUT_NORMAL)
2123		log_info("%s\n", fio_version_string);
2124
2125	return 0;
2126}
2127
2128void options_default_fill(struct thread_options *o)
2129{
2130	memcpy(o, &def_thread.o, sizeof(*o));
2131}
2132