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, size_t buf_size,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	buf[buf_size - 1] = '\0';
956	strncpy(buf, o->filename_format, buf_size - 1);
957
958	memset(copy, 0, sizeof(copy));
959	for (f = &fpre_keywords[0]; f->keyword; f++) {
960		do {
961			size_t pre_len, post_start = 0;
962			char *str, *dst = copy;
963
964			str = strcasestr(buf, f->keyword);
965			if (!str)
966				break;
967
968			pre_len = str - buf;
969			if (strlen(str) != f->strlen)
970				post_start = pre_len + f->strlen;
971
972			if (pre_len) {
973				strncpy(dst, buf, pre_len);
974				dst += pre_len;
975				dst_left -= pre_len;
976			}
977
978			switch (f->key) {
979			case FPRE_JOBNAME: {
980				int ret;
981
982				ret = snprintf(dst, dst_left, "%s", jobname);
983				if (ret < 0)
984					break;
985				dst += ret;
986				dst_left -= ret;
987				break;
988				}
989			case FPRE_JOBNUM: {
990				int ret;
991
992				ret = snprintf(dst, dst_left, "%d", jobnum);
993				if (ret < 0)
994					break;
995				dst += ret;
996				dst_left -= ret;
997				break;
998				}
999			case FPRE_FILENUM: {
1000				int ret;
1001
1002				ret = snprintf(dst, dst_left, "%d", filenum);
1003				if (ret < 0)
1004					break;
1005				dst += ret;
1006				dst_left -= ret;
1007				break;
1008				}
1009			default:
1010				assert(0);
1011				break;
1012			}
1013
1014			if (post_start)
1015				strncpy(dst, buf + post_start, dst_left);
1016
1017			strncpy(buf, copy, buf_size - 1);
1018		} while (1);
1019	}
1020
1021	return buf;
1022}
1023
1024int parse_dryrun(void)
1025{
1026	return dump_cmdline || parse_only;
1027}
1028
1029/*
1030 * Adds a job to the list of things todo. Sanitizes the various options
1031 * to make sure we don't have conflicts, and initializes various
1032 * members of td.
1033 */
1034static int add_job(struct thread_data *td, const char *jobname, int job_add_num,
1035		   int recursed, int client_type)
1036{
1037	unsigned int i;
1038	char fname[PATH_MAX];
1039	int numjobs, file_alloced;
1040	struct thread_options *o = &td->o;
1041
1042	/*
1043	 * the def_thread is just for options, it's not a real job
1044	 */
1045	if (td == &def_thread)
1046		return 0;
1047
1048	init_flags(td);
1049
1050	/*
1051	 * if we are just dumping the output command line, don't add the job
1052	 */
1053	if (parse_dryrun()) {
1054		put_job(td);
1055		return 0;
1056	}
1057
1058	td->client_type = client_type;
1059
1060	if (profile_td_init(td))
1061		goto err;
1062
1063	if (ioengine_load(td))
1064		goto err;
1065
1066	if (o->odirect)
1067		td->io_ops->flags |= FIO_RAWIO;
1068
1069	file_alloced = 0;
1070	if (!o->filename && !td->files_index && !o->read_iolog_file) {
1071		file_alloced = 1;
1072
1073		if (o->nr_files == 1 && exists_and_not_file(jobname))
1074			add_file(td, jobname, job_add_num, 0);
1075		else {
1076			for (i = 0; i < o->nr_files; i++)
1077				add_file(td, make_filename(fname, sizeof(fname), o, jobname, job_add_num, i), job_add_num, 0);
1078		}
1079	}
1080
1081	if (fixup_options(td))
1082		goto err;
1083
1084	flow_init_job(td);
1085
1086	/*
1087	 * IO engines only need this for option callbacks, and the address may
1088	 * change in subprocesses.
1089	 */
1090	if (td->eo)
1091		*(struct thread_data **)td->eo = NULL;
1092
1093	if (td->io_ops->flags & FIO_DISKLESSIO) {
1094		struct fio_file *f;
1095
1096		for_each_file(td, f, i)
1097			f->real_file_size = -1ULL;
1098	}
1099
1100	td->mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1101
1102	td->ts.clat_percentiles = o->clat_percentiles;
1103	td->ts.percentile_precision = o->percentile_precision;
1104	memcpy(td->ts.percentile_list, o->percentile_list, sizeof(o->percentile_list));
1105
1106	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1107		td->ts.clat_stat[i].min_val = ULONG_MAX;
1108		td->ts.slat_stat[i].min_val = ULONG_MAX;
1109		td->ts.lat_stat[i].min_val = ULONG_MAX;
1110		td->ts.bw_stat[i].min_val = ULONG_MAX;
1111	}
1112	td->ddir_seq_nr = o->ddir_seq_nr;
1113
1114	if ((o->stonewall || o->new_group) && prev_group_jobs) {
1115		prev_group_jobs = 0;
1116		groupid++;
1117	}
1118
1119	td->groupid = groupid;
1120	prev_group_jobs++;
1121
1122	if (setup_random_seeds(td)) {
1123		td_verror(td, errno, "init_random_state");
1124		goto err;
1125	}
1126
1127	if (setup_rate(td))
1128		goto err;
1129
1130	if (o->lat_log_file || write_lat_log) {
1131		setup_log(&td->lat_log, o->log_avg_msec, IO_LOG_TYPE_LAT);
1132		setup_log(&td->slat_log, o->log_avg_msec, IO_LOG_TYPE_SLAT);
1133		setup_log(&td->clat_log, o->log_avg_msec, IO_LOG_TYPE_CLAT);
1134	}
1135	if (o->bw_log_file || write_bw_log)
1136		setup_log(&td->bw_log, o->log_avg_msec, IO_LOG_TYPE_BW);
1137	if (o->iops_log_file)
1138		setup_log(&td->iops_log, o->log_avg_msec, IO_LOG_TYPE_IOPS);
1139
1140	if (!o->name)
1141		o->name = strdup(jobname);
1142
1143	if (output_format == FIO_OUTPUT_NORMAL) {
1144		if (!job_add_num) {
1145			if (is_backend && !recursed)
1146				fio_server_send_add_job(td);
1147
1148			if (!(td->io_ops->flags & FIO_NOIO)) {
1149				char *c1, *c2, *c3, *c4;
1150				char *c5 = NULL, *c6 = NULL;
1151
1152				c1 = fio_uint_to_kmg(o->min_bs[DDIR_READ]);
1153				c2 = fio_uint_to_kmg(o->max_bs[DDIR_READ]);
1154				c3 = fio_uint_to_kmg(o->min_bs[DDIR_WRITE]);
1155				c4 = fio_uint_to_kmg(o->max_bs[DDIR_WRITE]);
1156
1157				if (!o->bs_is_seq_rand) {
1158					c5 = fio_uint_to_kmg(o->min_bs[DDIR_TRIM]);
1159					c6 = fio_uint_to_kmg(o->max_bs[DDIR_TRIM]);
1160				}
1161
1162				log_info("%s: (g=%d): rw=%s, ", td->o.name,
1163							td->groupid,
1164							ddir_str(o->td_ddir));
1165
1166				if (o->bs_is_seq_rand)
1167					log_info("bs(seq/rand)=%s-%s/%s-%s, ",
1168							c1, c2, c3, c4);
1169				else
1170					log_info("bs=%s-%s/%s-%s/%s-%s, ",
1171							c1, c2, c3, c4, c5, c6);
1172
1173				log_info("ioengine=%s, iodepth=%u\n",
1174						td->io_ops->name, o->iodepth);
1175
1176				free(c1);
1177				free(c2);
1178				free(c3);
1179				free(c4);
1180				free(c5);
1181				free(c6);
1182			}
1183		} else if (job_add_num == 1)
1184			log_info("...\n");
1185	}
1186
1187	/*
1188	 * recurse add identical jobs, clear numjobs and stonewall options
1189	 * as they don't apply to sub-jobs
1190	 */
1191	numjobs = o->numjobs;
1192	while (--numjobs) {
1193		struct thread_data *td_new = get_new_job(0, td, 1);
1194
1195		if (!td_new)
1196			goto err;
1197
1198		td_new->o.numjobs = 1;
1199		td_new->o.stonewall = 0;
1200		td_new->o.new_group = 0;
1201
1202		if (file_alloced) {
1203			if (td_new->files) {
1204				struct fio_file *f;
1205				for_each_file(td_new, f, i) {
1206					if (f->file_name)
1207						sfree(f->file_name);
1208					sfree(f);
1209				}
1210				free(td_new->files);
1211				td_new->files = NULL;
1212			}
1213			td_new->files_index = 0;
1214			td_new->files_size = 0;
1215			if (td_new->o.filename) {
1216				free(td_new->o.filename);
1217				td_new->o.filename = NULL;
1218			}
1219		}
1220
1221		if (add_job(td_new, jobname, numjobs, 1, client_type))
1222			goto err;
1223	}
1224
1225	return 0;
1226err:
1227	put_job(td);
1228	return -1;
1229}
1230
1231/*
1232 * Parse as if 'o' was a command line
1233 */
1234void add_job_opts(const char **o, int client_type)
1235{
1236	struct thread_data *td, *td_parent;
1237	int i, in_global = 1;
1238	char jobname[32];
1239
1240	i = 0;
1241	td_parent = td = NULL;
1242	while (o[i]) {
1243		if (!strncmp(o[i], "name", 4)) {
1244			in_global = 0;
1245			if (td)
1246				add_job(td, jobname, 0, 0, client_type);
1247			td = NULL;
1248			sprintf(jobname, "%s", o[i] + 5);
1249		}
1250		if (in_global && !td_parent)
1251			td_parent = get_new_job(1, &def_thread, 0);
1252		else if (!in_global && !td) {
1253			if (!td_parent)
1254				td_parent = &def_thread;
1255			td = get_new_job(0, td_parent, 0);
1256		}
1257		if (in_global)
1258			fio_options_parse(td_parent, (char **) &o[i], 1, 0);
1259		else
1260			fio_options_parse(td, (char **) &o[i], 1, 0);
1261		i++;
1262	}
1263
1264	if (td)
1265		add_job(td, jobname, 0, 0, client_type);
1266}
1267
1268static int skip_this_section(const char *name)
1269{
1270	int i;
1271
1272	if (!nr_job_sections)
1273		return 0;
1274	if (!strncmp(name, "global", 6))
1275		return 0;
1276
1277	for (i = 0; i < nr_job_sections; i++)
1278		if (!strcmp(job_sections[i], name))
1279			return 0;
1280
1281	return 1;
1282}
1283
1284static int is_empty_or_comment(char *line)
1285{
1286	unsigned int i;
1287
1288	for (i = 0; i < strlen(line); i++) {
1289		if (line[i] == ';')
1290			return 1;
1291		if (line[i] == '#')
1292			return 1;
1293		if (!isspace((int) line[i]) && !iscntrl((int) line[i]))
1294			return 0;
1295	}
1296
1297	return 1;
1298}
1299
1300/*
1301 * This is our [ini] type file parser.
1302 */
1303int parse_jobs_ini(char *file, int is_buf, int stonewall_flag, int type)
1304{
1305	unsigned int global;
1306	struct thread_data *td;
1307	char *string, *name;
1308	FILE *f;
1309	char *p;
1310	int ret = 0, stonewall;
1311	int first_sect = 1;
1312	int skip_fgets = 0;
1313	int inside_skip = 0;
1314	char **opts;
1315	int i, alloc_opts, num_opts;
1316
1317	if (is_buf)
1318		f = NULL;
1319	else {
1320		if (!strcmp(file, "-"))
1321			f = stdin;
1322		else
1323			f = fopen(file, "r");
1324
1325		if (!f) {
1326			perror("fopen job file");
1327			return 1;
1328		}
1329	}
1330
1331	string = malloc(4096);
1332
1333	/*
1334	 * it's really 256 + small bit, 280 should suffice
1335	 */
1336	name = malloc(280);
1337	memset(name, 0, 280);
1338
1339	alloc_opts = 8;
1340	opts = malloc(sizeof(char *) * alloc_opts);
1341	num_opts = 0;
1342
1343	stonewall = stonewall_flag;
1344	do {
1345		/*
1346		 * if skip_fgets is set, we already have loaded a line we
1347		 * haven't handled.
1348		 */
1349		if (!skip_fgets) {
1350			if (is_buf)
1351				p = strsep(&file, "\n");
1352			else
1353				p = fgets(string, 4096, f);
1354			if (!p)
1355				break;
1356		}
1357
1358		skip_fgets = 0;
1359		strip_blank_front(&p);
1360		strip_blank_end(p);
1361
1362		if (is_empty_or_comment(p))
1363			continue;
1364		if (sscanf(p, "[%255[^\n]]", name) != 1) {
1365			if (inside_skip)
1366				continue;
1367			log_err("fio: option <%s> outside of [] job section\n",
1368									p);
1369			break;
1370		}
1371
1372		name[strlen(name) - 1] = '\0';
1373
1374		if (skip_this_section(name)) {
1375			inside_skip = 1;
1376			continue;
1377		} else
1378			inside_skip = 0;
1379
1380		global = !strncmp(name, "global", 6);
1381
1382		if (dump_cmdline) {
1383			if (first_sect)
1384				log_info("fio ");
1385			if (!global)
1386				log_info("--name=%s ", name);
1387			first_sect = 0;
1388		}
1389
1390		td = get_new_job(global, &def_thread, 0);
1391		if (!td) {
1392			ret = 1;
1393			break;
1394		}
1395
1396		/*
1397		 * Separate multiple job files by a stonewall
1398		 */
1399		if (!global && stonewall) {
1400			td->o.stonewall = stonewall;
1401			stonewall = 0;
1402		}
1403
1404		num_opts = 0;
1405		memset(opts, 0, alloc_opts * sizeof(char *));
1406
1407		while (1) {
1408			if (is_buf)
1409				p = strsep(&file, "\n");
1410			else
1411				p = fgets(string, 4096, f);
1412			if (!p)
1413				break;
1414
1415			if (is_empty_or_comment(p))
1416				continue;
1417
1418			strip_blank_front(&p);
1419
1420			/*
1421			 * new section, break out and make sure we don't
1422			 * fgets() a new line at the top.
1423			 */
1424			if (p[0] == '[') {
1425				skip_fgets = 1;
1426				break;
1427			}
1428
1429			strip_blank_end(p);
1430
1431			if (num_opts == alloc_opts) {
1432				alloc_opts <<= 1;
1433				opts = realloc(opts,
1434						alloc_opts * sizeof(char *));
1435			}
1436
1437			opts[num_opts] = strdup(p);
1438			num_opts++;
1439		}
1440
1441		ret = fio_options_parse(td, opts, num_opts, dump_cmdline);
1442		if (!ret)
1443			ret = add_job(td, name, 0, 0, type);
1444		else {
1445			log_err("fio: job %s dropped\n", name);
1446			put_job(td);
1447		}
1448
1449		for (i = 0; i < num_opts; i++)
1450			free(opts[i]);
1451		num_opts = 0;
1452	} while (!ret);
1453
1454	if (dump_cmdline)
1455		log_info("\n");
1456
1457	i = 0;
1458	while (i < nr_job_sections) {
1459		free(job_sections[i]);
1460		i++;
1461	}
1462
1463	free(string);
1464	free(name);
1465	free(opts);
1466	if (!is_buf && f != stdin)
1467		fclose(f);
1468	return ret;
1469}
1470
1471static int fill_def_thread(void)
1472{
1473	memset(&def_thread, 0, sizeof(def_thread));
1474
1475	fio_getaffinity(getpid(), &def_thread.o.cpumask);
1476	def_thread.o.error_dump = 1;
1477
1478	/*
1479	 * fill default options
1480	 */
1481	fio_fill_default_options(&def_thread);
1482	return 0;
1483}
1484
1485static void usage(const char *name)
1486{
1487	printf("%s\n", fio_version_string);
1488	printf("%s [options] [job options] <job file(s)>\n", name);
1489	printf("  --debug=options\tEnable debug logging. May be one/more of:\n"
1490		"\t\t\tprocess,file,io,mem,blktrace,verify,random,parse,\n"
1491		"\t\t\tdiskutil,job,mutex,profile,time,net,rate\n");
1492	printf("  --parse-only\t\tParse options only, don't start any IO\n");
1493	printf("  --output\t\tWrite output to file\n");
1494	printf("  --runtime\t\tRuntime in seconds\n");
1495	printf("  --latency-log\t\tGenerate per-job latency logs\n");
1496	printf("  --bandwidth-log\tGenerate per-job bandwidth logs\n");
1497	printf("  --minimal\t\tMinimal (terse) output\n");
1498	printf("  --output-format=x\tOutput format (terse,json,normal)\n");
1499	printf("  --terse-version=x\tSet terse version output format to 'x'\n");
1500	printf("  --version\t\tPrint version info and exit\n");
1501	printf("  --help\t\tPrint this page\n");
1502	printf("  --cpuclock-test\tPerform test/validation of CPU clock\n");
1503	printf("  --crctest\t\tTest speed of checksum functions\n");
1504	printf("  --cmdhelp=cmd\t\tPrint command help, \"all\" for all of"
1505		" them\n");
1506	printf("  --enghelp=engine\tPrint ioengine help, or list"
1507		" available ioengines\n");
1508	printf("  --enghelp=engine,cmd\tPrint help for an ioengine"
1509		" cmd\n");
1510	printf("  --showcmd\t\tTurn a job file into command line options\n");
1511	printf("  --eta=when\t\tWhen ETA estimate should be printed\n");
1512	printf("            \t\tMay be \"always\", \"never\" or \"auto\"\n");
1513	printf("  --eta-newline=time\tForce a new line for every 'time'");
1514	printf(" period passed\n");
1515	printf("  --status-interval=t\tForce full status dump every");
1516	printf(" 't' period passed\n");
1517	printf("  --readonly\t\tTurn on safety read-only checks, preventing"
1518		" writes\n");
1519	printf("  --section=name\tOnly run specified section in job file\n");
1520	printf("  --alloc-size=kb\tSet smalloc pool to this size in kb"
1521		" (def 1024)\n");
1522	printf("  --warnings-fatal\tFio parser warnings are fatal\n");
1523	printf("  --max-jobs=nr\t\tMaximum number of threads/processes to support\n");
1524	printf("  --server=args\t\tStart a backend fio server\n");
1525	printf("  --daemonize=pidfile\tBackground fio server, write pid to file\n");
1526	printf("  --client=hostname\tTalk to remote backend fio server at hostname\n");
1527	printf("  --idle-prof=option\tReport cpu idleness on a system or percpu basis\n"
1528		"\t\t\t(option=system,percpu) or run unit work\n"
1529		"\t\t\tcalibration only (option=calibrate)\n");
1530	printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
1531	printf("\n                   Jens Axboe <jaxboe@fusionio.com>");
1532	printf("\n                   Jens Axboe <axboe@fb.com>\n");
1533}
1534
1535#ifdef FIO_INC_DEBUG
1536struct debug_level debug_levels[] = {
1537	{ .name = "process",
1538	  .help = "Process creation/exit logging",
1539	  .shift = FD_PROCESS,
1540	},
1541	{ .name = "file",
1542	  .help = "File related action logging",
1543	  .shift = FD_FILE,
1544	},
1545	{ .name = "io",
1546	  .help = "IO and IO engine action logging (offsets, queue, completions, etc)",
1547	  .shift = FD_IO,
1548	},
1549	{ .name = "mem",
1550	  .help = "Memory allocation/freeing logging",
1551	  .shift = FD_MEM,
1552	},
1553	{ .name = "blktrace",
1554	  .help = "blktrace action logging",
1555	  .shift = FD_BLKTRACE,
1556	},
1557	{ .name = "verify",
1558	  .help = "IO verification action logging",
1559	  .shift = FD_VERIFY,
1560	},
1561	{ .name = "random",
1562	  .help = "Random generation logging",
1563	  .shift = FD_RANDOM,
1564	},
1565	{ .name = "parse",
1566	  .help = "Parser logging",
1567	  .shift = FD_PARSE,
1568	},
1569	{ .name = "diskutil",
1570	  .help = "Disk utility logging actions",
1571	  .shift = FD_DISKUTIL,
1572	},
1573	{ .name = "job",
1574	  .help = "Logging related to creating/destroying jobs",
1575	  .shift = FD_JOB,
1576	},
1577	{ .name = "mutex",
1578	  .help = "Mutex logging",
1579	  .shift = FD_MUTEX
1580	},
1581	{ .name	= "profile",
1582	  .help = "Logging related to profiles",
1583	  .shift = FD_PROFILE,
1584	},
1585	{ .name = "time",
1586	  .help = "Logging related to time keeping functions",
1587	  .shift = FD_TIME,
1588	},
1589	{ .name = "net",
1590	  .help = "Network logging",
1591	  .shift = FD_NET,
1592	},
1593	{ .name = "rate",
1594	  .help = "Rate logging",
1595	  .shift = FD_RATE,
1596	},
1597	{ .name = NULL, },
1598};
1599
1600static int set_debug(const char *string)
1601{
1602	struct debug_level *dl;
1603	char *p = (char *) string;
1604	char *opt;
1605	int i;
1606
1607	if (!strcmp(string, "?") || !strcmp(string, "help")) {
1608		log_info("fio: dumping debug options:");
1609		for (i = 0; debug_levels[i].name; i++) {
1610			dl = &debug_levels[i];
1611			log_info("%s,", dl->name);
1612		}
1613		log_info("all\n");
1614		return 1;
1615	}
1616
1617	while ((opt = strsep(&p, ",")) != NULL) {
1618		int found = 0;
1619
1620		if (!strncmp(opt, "all", 3)) {
1621			log_info("fio: set all debug options\n");
1622			fio_debug = ~0UL;
1623			continue;
1624		}
1625
1626		for (i = 0; debug_levels[i].name; i++) {
1627			dl = &debug_levels[i];
1628			found = !strncmp(opt, dl->name, strlen(dl->name));
1629			if (!found)
1630				continue;
1631
1632			if (dl->shift == FD_JOB) {
1633				opt = strchr(opt, ':');
1634				if (!opt) {
1635					log_err("fio: missing job number\n");
1636					break;
1637				}
1638				opt++;
1639				fio_debug_jobno = atoi(opt);
1640				log_info("fio: set debug jobno %d\n",
1641							fio_debug_jobno);
1642			} else {
1643				log_info("fio: set debug option %s\n", opt);
1644				fio_debug |= (1UL << dl->shift);
1645			}
1646			break;
1647		}
1648
1649		if (!found)
1650			log_err("fio: debug mask %s not found\n", opt);
1651	}
1652	return 0;
1653}
1654#else
1655static int set_debug(const char *string)
1656{
1657	log_err("fio: debug tracing not included in build\n");
1658	return 1;
1659}
1660#endif
1661
1662static void fio_options_fill_optstring(void)
1663{
1664	char *ostr = cmd_optstr;
1665	int i, c;
1666
1667	c = i = 0;
1668	while (l_opts[i].name) {
1669		ostr[c++] = l_opts[i].val;
1670		if (l_opts[i].has_arg == required_argument)
1671			ostr[c++] = ':';
1672		else if (l_opts[i].has_arg == optional_argument) {
1673			ostr[c++] = ':';
1674			ostr[c++] = ':';
1675		}
1676		i++;
1677	}
1678	ostr[c] = '\0';
1679}
1680
1681static int client_flag_set(char c)
1682{
1683	int i;
1684
1685	i = 0;
1686	while (l_opts[i].name) {
1687		int val = l_opts[i].val;
1688
1689		if (c == (val & 0xff))
1690			return (val & FIO_CLIENT_FLAG);
1691
1692		i++;
1693	}
1694
1695	return 0;
1696}
1697
1698static void parse_cmd_client(void *client, char *opt)
1699{
1700	fio_client_add_cmd_option(client, opt);
1701}
1702
1703int parse_cmd_line(int argc, char *argv[], int client_type)
1704{
1705	struct thread_data *td = NULL;
1706	int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1707	char *ostr = cmd_optstr;
1708	void *pid_file = NULL;
1709	void *cur_client = NULL;
1710	int backend = 0;
1711
1712	/*
1713	 * Reset optind handling, since we may call this multiple times
1714	 * for the backend.
1715	 */
1716	optind = 1;
1717
1718	while ((c = getopt_long_only(argc, argv, ostr, l_opts, &lidx)) != -1) {
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			did_arg = 1;
1776			if (!cur_client) {
1777				usage(argv[0]);
1778				do_exit++;
1779			}
1780			break;
1781		case 'c':
1782			did_arg = 1;
1783			if (!cur_client) {
1784				fio_show_option_help(optarg);
1785				do_exit++;
1786			}
1787			break;
1788		case 'i':
1789			did_arg = 1;
1790			if (!cur_client) {
1791				fio_show_ioengine_help(optarg);
1792				do_exit++;
1793			}
1794			break;
1795		case 's':
1796			did_arg = 1;
1797			dump_cmdline = 1;
1798			break;
1799		case 'r':
1800			read_only = 1;
1801			break;
1802		case 'v':
1803			did_arg = 1;
1804			if (!cur_client) {
1805				log_info("%s\n", fio_version_string);
1806				do_exit++;
1807			}
1808			break;
1809		case 'V':
1810			terse_version = atoi(optarg);
1811			if (!(terse_version == 2 || terse_version == 3 ||
1812			     terse_version == 4)) {
1813				log_err("fio: bad terse version format\n");
1814				exit_val = 1;
1815				do_exit++;
1816			}
1817			break;
1818		case 'e':
1819			if (!strcmp("always", optarg))
1820				eta_print = FIO_ETA_ALWAYS;
1821			else if (!strcmp("never", optarg))
1822				eta_print = FIO_ETA_NEVER;
1823			break;
1824		case 'E': {
1825			long long t = 0;
1826
1827			if (str_to_decimal(optarg, &t, 0, NULL, 1)) {
1828				log_err("fio: failed parsing eta time %s\n", optarg);
1829				exit_val = 1;
1830				do_exit++;
1831			}
1832			eta_new_line = t;
1833			break;
1834			}
1835		case 'd':
1836			if (set_debug(optarg))
1837				do_exit++;
1838			break;
1839		case 'P':
1840			did_arg = 1;
1841			parse_only = 1;
1842			break;
1843		case 'x': {
1844			size_t new_size;
1845
1846			if (!strcmp(optarg, "global")) {
1847				log_err("fio: can't use global as only "
1848					"section\n");
1849				do_exit++;
1850				exit_val = 1;
1851				break;
1852			}
1853			new_size = (nr_job_sections + 1) * sizeof(char *);
1854			job_sections = realloc(job_sections, new_size);
1855			job_sections[nr_job_sections] = strdup(optarg);
1856			nr_job_sections++;
1857			break;
1858			}
1859		case 'p':
1860			did_arg = 1;
1861			if (exec_profile)
1862				free(exec_profile);
1863			exec_profile = strdup(optarg);
1864			break;
1865		case FIO_GETOPT_JOB: {
1866			const char *opt = l_opts[lidx].name;
1867			char *val = optarg;
1868
1869			if (!strncmp(opt, "name", 4) && td) {
1870				ret = add_job(td, td->o.name ?: "fio", 0, 0, client_type);
1871				if (ret)
1872					goto out_free;
1873				td = NULL;
1874				did_arg = 1;
1875			}
1876			if (!td) {
1877				int is_section = !strncmp(opt, "name", 4);
1878				int global = 0;
1879
1880				if (!is_section || !strncmp(val, "global", 6))
1881					global = 1;
1882
1883				if (is_section && skip_this_section(val))
1884					continue;
1885
1886				td = get_new_job(global, &def_thread, 1);
1887				if (!td || ioengine_load(td))
1888					goto out_free;
1889				fio_options_set_ioengine_opts(l_opts, td);
1890			}
1891
1892			if ((!val || !strlen(val)) &&
1893			    l_opts[lidx].has_arg == required_argument) {
1894				log_err("fio: option %s requires an argument\n", opt);
1895				ret = 1;
1896			} else
1897				ret = fio_cmd_option_parse(td, opt, val);
1898
1899			if (ret) {
1900				if (td) {
1901					put_job(td);
1902					td = NULL;
1903				}
1904				do_exit++;
1905			}
1906
1907			if (!ret && !strcmp(opt, "ioengine")) {
1908				free_ioengine(td);
1909				if (ioengine_load(td))
1910					goto out_free;
1911				fio_options_set_ioengine_opts(l_opts, td);
1912			}
1913			break;
1914		}
1915		case FIO_GETOPT_IOENGINE: {
1916			const char *opt = l_opts[lidx].name;
1917			char *val = optarg;
1918
1919			if (!td)
1920				break;
1921
1922			ret = fio_cmd_ioengine_option_parse(td, opt, val);
1923			break;
1924		}
1925		case 'w':
1926			warnings_fatal = 1;
1927			break;
1928		case 'j':
1929			max_jobs = atoi(optarg);
1930			if (!max_jobs || max_jobs > REAL_MAX_JOBS) {
1931				log_err("fio: invalid max jobs: %d\n", max_jobs);
1932				do_exit++;
1933				exit_val = 1;
1934			}
1935			break;
1936		case 'S':
1937			did_arg = 1;
1938			if (nr_clients) {
1939				log_err("fio: can't be both client and server\n");
1940				do_exit++;
1941				exit_val = 1;
1942				break;
1943			}
1944			if (optarg)
1945				fio_server_set_arg(optarg);
1946			is_backend = 1;
1947			backend = 1;
1948			break;
1949		case 'D':
1950			if (pid_file)
1951				free(pid_file);
1952			pid_file = strdup(optarg);
1953			break;
1954		case 'I':
1955			if ((ret = fio_idle_prof_parse_opt(optarg))) {
1956				/* exit on error and calibration only */
1957				did_arg = 1;
1958				do_exit++;
1959				if (ret == -1)
1960					exit_val = 1;
1961			}
1962			break;
1963		case 'C':
1964			did_arg = 1;
1965			if (is_backend) {
1966				log_err("fio: can't be both client and server\n");
1967				do_exit++;
1968				exit_val = 1;
1969				break;
1970			}
1971			if (fio_client_add(&fio_client_ops, optarg, &cur_client)) {
1972				log_err("fio: failed adding client %s\n", optarg);
1973				do_exit++;
1974				exit_val = 1;
1975				break;
1976			}
1977			/*
1978			 * If the next argument exists and isn't an option,
1979			 * assume it's a job file for this client only.
1980			 */
1981			while (optind < argc) {
1982				if (!strncmp(argv[optind], "--", 2) ||
1983				    !strncmp(argv[optind], "-", 1))
1984					break;
1985
1986				fio_client_add_ini_file(cur_client, argv[optind]);
1987				optind++;
1988			}
1989			break;
1990		case 'T':
1991			did_arg = 1;
1992			do_exit++;
1993			exit_val = fio_monotonic_clocktest();
1994			break;
1995		case 'G':
1996			did_arg = 1;
1997			do_exit++;
1998			exit_val = fio_crctest(optarg);
1999			break;
2000		case 'L': {
2001			long long val;
2002
2003			if (check_str_time(optarg, &val, 0)) {
2004				log_err("fio: failed parsing time %s\n", optarg);
2005				do_exit++;
2006				exit_val = 1;
2007				break;
2008			}
2009			status_interval = val * 1000;
2010			break;
2011			}
2012		case '?':
2013			log_err("%s: unrecognized option '%s'\n", argv[0],
2014							argv[optind - 1]);
2015		default:
2016			do_exit++;
2017			exit_val = 1;
2018			break;
2019		}
2020		if (do_exit)
2021			break;
2022	}
2023
2024	if (do_exit && !(is_backend || nr_clients))
2025		exit(exit_val);
2026
2027	if (nr_clients && fio_clients_connect())
2028		exit(1);
2029
2030	if (is_backend && backend)
2031		return fio_start_server(pid_file);
2032	else if (pid_file)
2033		free(pid_file);
2034
2035	if (td) {
2036		if (!ret) {
2037			ret = add_job(td, td->o.name ?: "fio", 0, 0, client_type);
2038			if (ret)
2039				did_arg = 1;
2040		}
2041	}
2042
2043	while (!ret && optind < argc) {
2044		ini_idx++;
2045		ini_file = realloc(ini_file, ini_idx * sizeof(char *));
2046		ini_file[ini_idx - 1] = strdup(argv[optind]);
2047		optind++;
2048	}
2049
2050out_free:
2051	if (pid_file)
2052		free(pid_file);
2053
2054	return ini_idx;
2055}
2056
2057int fio_init_options(void)
2058{
2059	f_out = stdout;
2060	f_err = stderr;
2061
2062	fio_options_fill_optstring();
2063	fio_options_dup_and_init(l_opts);
2064
2065	atexit(free_shm);
2066
2067	if (fill_def_thread())
2068		return 1;
2069
2070	return 0;
2071}
2072
2073extern int fio_check_options(struct thread_options *);
2074
2075int parse_options(int argc, char *argv[])
2076{
2077	const int type = FIO_CLIENT_TYPE_CLI;
2078	int job_files, i;
2079
2080	if (fio_init_options())
2081		return 1;
2082	if (fio_test_cconv(&def_thread.o))
2083		log_err("fio: failed internal cconv test\n");
2084
2085	job_files = parse_cmd_line(argc, argv, type);
2086
2087	if (job_files > 0) {
2088		for (i = 0; i < job_files; i++) {
2089			if (i && fill_def_thread())
2090				return 1;
2091			if (nr_clients) {
2092				if (fio_clients_send_ini(ini_file[i]))
2093					return 1;
2094				free(ini_file[i]);
2095			} else if (!is_backend) {
2096				if (parse_jobs_ini(ini_file[i], 0, i, type))
2097					return 1;
2098				free(ini_file[i]);
2099			}
2100		}
2101	} else if (nr_clients) {
2102		if (fill_def_thread())
2103			return 1;
2104		if (fio_clients_send_ini(NULL))
2105			return 1;
2106	}
2107
2108	free(ini_file);
2109	fio_options_free(&def_thread);
2110	filesetup_mem_free();
2111
2112	if (!thread_number) {
2113		if (parse_dryrun())
2114			return 0;
2115		if (exec_profile)
2116			return 0;
2117		if (is_backend || nr_clients)
2118			return 0;
2119		if (did_arg)
2120			return 0;
2121
2122		log_err("No jobs(s) defined\n\n");
2123
2124		if (!did_arg) {
2125			usage(argv[0]);
2126			return 1;
2127		}
2128
2129		return 0;
2130	}
2131
2132	if (def_thread.o.gtod_offload) {
2133		fio_gtod_init();
2134		fio_gtod_offload = 1;
2135		fio_gtod_cpu = def_thread.o.gtod_cpu;
2136	}
2137
2138	if (output_format == FIO_OUTPUT_NORMAL)
2139		log_info("%s\n", fio_version_string);
2140
2141	return 0;
2142}
2143
2144void options_default_fill(struct thread_options *o)
2145{
2146	memcpy(o, &def_thread.o, sizeof(*o));
2147}
2148