init.c revision 58c55ba02aca24ed30beb343fcb1dc3df4667389
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 <getopt.h>
12#include <sys/ipc.h>
13#include <sys/shm.h>
14#include <sys/types.h>
15#include <sys/stat.h>
16
17#include "fio.h"
18#include "parse.h"
19#include "smalloc.h"
20#include "filehash.h"
21#include "verify.h"
22#include "profile.h"
23
24static char fio_version_string[] = "fio 1.37";
25
26#define FIO_RANDSEED		(0xb1899bedUL)
27
28static char **ini_file;
29static int max_jobs = MAX_JOBS;
30static int dump_cmdline;
31
32static struct thread_data def_thread;
33struct thread_data *threads = NULL;
34
35int exitall_on_terminate = 0;
36int terse_output = 0;
37int eta_print;
38unsigned long long mlock_size = 0;
39FILE *f_out = NULL;
40FILE *f_err = NULL;
41char *job_section = NULL;
42char *exec_profile = NULL;
43
44int write_bw_log = 0;
45int read_only = 0;
46
47static int def_timeout;
48static int write_lat_log;
49
50static int prev_group_jobs;
51
52unsigned long fio_debug = 0;
53unsigned int fio_debug_jobno = -1;
54unsigned int *fio_debug_jobp = NULL;
55
56/*
57 * Command line options. These will contain the above, plus a few
58 * extra that only pertain to fio itself and not jobs.
59 */
60static struct option l_opts[FIO_NR_OPTIONS] = {
61	{
62		.name		= "output",
63		.has_arg	= required_argument,
64		.val		= 'o',
65	},
66	{
67		.name		= "timeout",
68		.has_arg	= required_argument,
69		.val		= 't',
70	},
71	{
72		.name		= "latency-log",
73		.has_arg	= required_argument,
74		.val		= 'l',
75	},
76	{
77		.name		= "bandwidth-log",
78		.has_arg	= required_argument,
79		.val		= 'b',
80	},
81	{
82		.name		= "minimal",
83		.has_arg	= optional_argument,
84		.val		= 'm',
85	},
86	{
87		.name		= "version",
88		.has_arg	= no_argument,
89		.val		= 'v',
90	},
91	{
92		.name		= "help",
93		.has_arg	= no_argument,
94		.val		= 'h',
95	},
96	{
97		.name		= "cmdhelp",
98		.has_arg	= optional_argument,
99		.val		= 'c',
100	},
101	{
102		.name		= "showcmd",
103		.has_arg	= no_argument,
104		.val		= 's',
105	},
106	{
107		.name		= "readonly",
108		.has_arg	= no_argument,
109		.val		= 'r',
110	},
111	{
112		.name		= "eta",
113		.has_arg	= required_argument,
114		.val		= 'e',
115	},
116	{
117		.name		= "debug",
118		.has_arg	= required_argument,
119		.val		= 'd',
120	},
121	{
122		.name		= "section",
123		.has_arg	= required_argument,
124		.val		= 'x',
125	},
126	{
127		.name		= "alloc-size",
128		.has_arg	= required_argument,
129		.val		= 'a',
130	},
131	{
132		.name		= "profile",
133		.has_arg	= required_argument,
134		.val		= 'p',
135	},
136	{
137		.name		= NULL,
138	},
139};
140
141FILE *get_f_out()
142{
143	return f_out;
144}
145
146FILE *get_f_err()
147{
148	return f_err;
149}
150
151/*
152 * Return a free job structure.
153 */
154static struct thread_data *get_new_job(int global, struct thread_data *parent)
155{
156	struct thread_data *td;
157
158	if (global)
159		return &def_thread;
160	if (thread_number >= max_jobs)
161		return NULL;
162
163	td = &threads[thread_number++];
164	*td = *parent;
165
166	td->o.uid = td->o.gid = -1U;
167
168	dup_files(td, parent);
169	options_mem_dupe(td);
170
171	profile_add_hooks(td);
172
173	td->thread_number = thread_number;
174	return td;
175}
176
177static void put_job(struct thread_data *td)
178{
179	if (td == &def_thread)
180		return;
181
182	profile_td_exit(td);
183
184	if (td->error)
185		log_info("fio: %s\n", td->verror);
186
187	memset(&threads[td->thread_number - 1], 0, sizeof(*td));
188	thread_number--;
189}
190
191static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
192{
193	unsigned int bs = td->o.min_bs[ddir];
194	unsigned long long bytes_per_sec;
195
196	if (td->o.rate[ddir])
197		bytes_per_sec = td->o.rate[ddir];
198	else
199		bytes_per_sec = td->o.rate_iops[ddir] * bs;
200
201	if (!bytes_per_sec) {
202		log_err("rate lower than supported\n");
203		return -1;
204	}
205
206	td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
207	td->rate_pending_usleep[ddir] = 0;
208	return 0;
209}
210
211static int setup_rate(struct thread_data *td)
212{
213	int ret = 0;
214
215	if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
216		ret = __setup_rate(td, DDIR_READ);
217	if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
218		ret |= __setup_rate(td, DDIR_WRITE);
219
220	return ret;
221}
222
223static int fixed_block_size(struct thread_options *o)
224{
225	return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
226		o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
227		o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
228}
229
230/*
231 * Lazy way of fixing up options that depend on each other. We could also
232 * define option callback handlers, but this is easier.
233 */
234static int fixup_options(struct thread_data *td)
235{
236	struct thread_options *o = &td->o;
237
238#ifndef FIO_HAVE_PSHARED_MUTEX
239	if (!td->o.use_thread) {
240		log_info("fio: this platform does not support process shared"
241			 " mutexes, forcing use of threads. Use the 'thread'"
242			 " option to get rid of this warning.\n");
243		td->o.use_thread = 1;
244	}
245#endif
246
247	if (o->write_iolog_file && o->read_iolog_file) {
248		log_err("fio: read iolog overrides write_iolog\n");
249		free(o->write_iolog_file);
250		o->write_iolog_file = NULL;
251	}
252
253	/*
254	 * only really works for sequential io for now, and with 1 file
255	 */
256	if (o->zone_size && td_random(td) && o->open_files == 1)
257		o->zone_size = 0;
258
259	/*
260	 * Reads can do overwrites, we always need to pre-create the file
261	 */
262	if (td_read(td) || td_rw(td))
263		o->overwrite = 1;
264
265	if (!o->min_bs[DDIR_READ])
266		o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
267	if (!o->max_bs[DDIR_READ])
268		o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
269	if (!o->min_bs[DDIR_WRITE])
270		o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
271	if (!o->max_bs[DDIR_WRITE])
272		o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
273
274	o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
275
276	/*
277	 * For random IO, allow blockalign offset other than min_bs.
278	 */
279	if (!o->ba[DDIR_READ] || !td_random(td))
280		o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
281	if (!o->ba[DDIR_WRITE] || !td_random(td))
282		o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
283
284	if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
285	    o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
286	    !td->o.norandommap) {
287		log_err("fio: Any use of blockalign= turns off randommap\n");
288		td->o.norandommap = 1;
289	}
290
291	if (!o->file_size_high)
292		o->file_size_high = o->file_size_low;
293
294	if (o->norandommap && o->verify != VERIFY_NONE
295	    && !fixed_block_size(o))  {
296		log_err("fio: norandommap given for variable block sizes, "
297			"verify disabled\n");
298		o->verify = VERIFY_NONE;
299	}
300	if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
301		log_err("fio: bs_unaligned may not work with raw io\n");
302
303	/*
304	 * thinktime_spin must be less than thinktime
305	 */
306	if (o->thinktime_spin > o->thinktime)
307		o->thinktime_spin = o->thinktime;
308
309	/*
310	 * The low water mark cannot be bigger than the iodepth
311	 */
312	if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
313		/*
314		 * syslet work around - if the workload is sequential,
315		 * we want to let the queue drain all the way down to
316		 * avoid seeking between async threads
317		 */
318		if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
319			o->iodepth_low = 1;
320		else
321			o->iodepth_low = o->iodepth;
322	}
323
324	/*
325	 * If batch number isn't set, default to the same as iodepth
326	 */
327	if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
328		o->iodepth_batch = o->iodepth;
329
330	if (o->nr_files > td->files_index)
331		o->nr_files = td->files_index;
332
333	if (o->open_files > o->nr_files || !o->open_files)
334		o->open_files = o->nr_files;
335
336	if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
337	    ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
338		o->rate_iops_min[1]))) {
339		log_err("fio: rate and rate_iops are mutually exclusive\n");
340		return 1;
341	}
342	if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
343	    (o->rate_iops[0] < o->rate_iops_min[0]) ||
344	    (o->rate_iops[1] < o->rate_iops_min[1])) {
345		log_err("fio: minimum rate exceeds rate\n");
346		return 1;
347	}
348
349	if (!o->timeout && o->time_based) {
350		log_err("fio: time_based requires a runtime/timeout setting\n");
351		o->time_based = 0;
352	}
353
354	if (o->fill_device && !o->size)
355		o->size = -1ULL;
356
357	if (td_rw(td) && td->o.verify != VERIFY_NONE)
358		log_info("fio: mixed read/write workload with verify. May not "
359		 "work as expected, unless you pre-populated the file\n");
360
361	if (td->o.verify != VERIFY_NONE)
362		td->o.refill_buffers = 1;
363
364	if (td->o.pre_read) {
365		td->o.invalidate_cache = 0;
366		if (td->io_ops->flags & FIO_PIPEIO)
367			log_info("fio: cannot pre-read files with an IO engine"
368				 " that isn't seekable. Pre-read disabled.\n");
369	}
370
371#ifndef FIO_HAVE_FDATASYNC
372	if (td->o.fdatasync_blocks) {
373		log_info("fio: this platform does not support fdatasync()"
374			 " falling back to using fsync().  Use the 'fsync'"
375			 " option instead of 'fdatasync' to get rid of"
376			 " this warning\n");
377		td->o.fsync_blocks = td->o.fdatasync_blocks;
378		td->o.fdatasync_blocks = 0;
379	}
380#endif
381
382	return 0;
383}
384
385/*
386 * This function leaks the buffer
387 */
388static char *to_kmg(unsigned int val)
389{
390	char *buf = malloc(32);
391	char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
392	char *p = post;
393
394	do {
395		if (val & 1023)
396			break;
397
398		val >>= 10;
399		p++;
400	} while (*p);
401
402	snprintf(buf, 31, "%u%c", val, *p);
403	return buf;
404}
405
406/* External engines are specified by "external:name.o") */
407static const char *get_engine_name(const char *str)
408{
409	char *p = strstr(str, ":");
410
411	if (!p)
412		return str;
413
414	p++;
415	strip_blank_front(&p);
416	strip_blank_end(p);
417	return p;
418}
419
420static int exists_and_not_file(const char *filename)
421{
422	struct stat sb;
423
424	if (lstat(filename, &sb) == -1)
425		return 0;
426
427	if (S_ISREG(sb.st_mode))
428		return 0;
429
430	return 1;
431}
432
433void td_fill_rand_seeds(struct thread_data *td)
434{
435	os_random_seed(td->rand_seeds[0], &td->bsrange_state);
436	os_random_seed(td->rand_seeds[1], &td->verify_state);
437	os_random_seed(td->rand_seeds[2], &td->rwmix_state);
438
439	if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
440		os_random_seed(td->rand_seeds[3], &td->next_file_state);
441
442	os_random_seed(td->rand_seeds[5], &td->file_size_state);
443
444	if (!td_random(td))
445		return;
446
447	if (td->o.rand_repeatable)
448		td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
449
450	os_random_seed(td->rand_seeds[4], &td->random_state);
451}
452
453/*
454 * Initialize the various random states we need (random io, block size ranges,
455 * read/write mix, etc).
456 */
457static int init_random_state(struct thread_data *td)
458{
459	int fd;
460
461	fd = open("/dev/urandom", O_RDONLY);
462	if (fd == -1) {
463		td_verror(td, errno, "open");
464		return 1;
465	}
466
467	if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
468	    (int) sizeof(td->rand_seeds)) {
469		td_verror(td, EIO, "read");
470		close(fd);
471		return 1;
472	}
473
474	close(fd);
475	td_fill_rand_seeds(td);
476	return 0;
477}
478
479/*
480 * Adds a job to the list of things todo. Sanitizes the various options
481 * to make sure we don't have conflicts, and initializes various
482 * members of td.
483 */
484static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
485{
486	const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
487				   "randread", "randwrite", "randrw" };
488	unsigned int i;
489	const char *engine;
490	char fname[PATH_MAX];
491	int numjobs, file_alloced;
492
493	/*
494	 * the def_thread is just for options, it's not a real job
495	 */
496	if (td == &def_thread)
497		return 0;
498
499	/*
500	 * if we are just dumping the output command line, don't add the job
501	 */
502	if (dump_cmdline) {
503		put_job(td);
504		return 0;
505	}
506
507	if (profile_td_init(td))
508		return 1;
509
510	engine = get_engine_name(td->o.ioengine);
511	td->io_ops = load_ioengine(td, engine);
512	if (!td->io_ops) {
513		log_err("fio: failed to load engine %s\n", engine);
514		goto err;
515	}
516
517	if (td->o.use_thread)
518		nr_thread++;
519	else
520		nr_process++;
521
522	if (td->o.odirect)
523		td->io_ops->flags |= FIO_RAWIO;
524
525	file_alloced = 0;
526	if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
527		file_alloced = 1;
528
529		if (td->o.nr_files == 1 && exists_and_not_file(jobname))
530			add_file(td, jobname);
531		else {
532			for (i = 0; i < td->o.nr_files; i++) {
533				sprintf(fname, "%s.%d.%d", jobname,
534							td->thread_number, i);
535				add_file(td, fname);
536			}
537		}
538	}
539
540	if (fixup_options(td))
541		goto err;
542
543	if (td->io_ops->flags & FIO_DISKLESSIO) {
544		struct fio_file *f;
545
546		for_each_file(td, f, i)
547			f->real_file_size = -1ULL;
548	}
549
550	td->mutex = fio_mutex_init(0);
551
552	td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
553	td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
554	td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
555	td->ddir_nr = td->o.ddir_nr;
556
557	if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
558		prev_group_jobs = 0;
559		groupid++;
560	}
561
562	td->groupid = groupid;
563	prev_group_jobs++;
564
565	if (init_random_state(td))
566		goto err;
567
568	if (setup_rate(td))
569		goto err;
570
571	if (td->o.write_lat_log) {
572		setup_log(&td->ts.slat_log);
573		setup_log(&td->ts.clat_log);
574	}
575	if (td->o.write_bw_log)
576		setup_log(&td->ts.bw_log);
577
578	if (!td->o.name)
579		td->o.name = strdup(jobname);
580
581	if (!terse_output) {
582		if (!job_add_num) {
583			if (!strcmp(td->io_ops->name, "cpuio")) {
584				log_info("%s: ioengine=cpu, cpuload=%u,"
585					 " cpucycle=%u\n", td->o.name,
586							td->o.cpuload,
587							td->o.cpucycle);
588			} else {
589				char *c1, *c2, *c3, *c4;
590
591				c1 = to_kmg(td->o.min_bs[DDIR_READ]);
592				c2 = to_kmg(td->o.max_bs[DDIR_READ]);
593				c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
594				c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
595
596				log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
597					 " ioengine=%s, iodepth=%u\n",
598						td->o.name, td->groupid,
599						ddir_str[td->o.td_ddir],
600						c1, c2, c3, c4,
601						td->io_ops->name,
602						td->o.iodepth);
603
604				free(c1);
605				free(c2);
606				free(c3);
607				free(c4);
608			}
609		} else if (job_add_num == 1)
610			log_info("...\n");
611	}
612
613	/*
614	 * recurse add identical jobs, clear numjobs and stonewall options
615	 * as they don't apply to sub-jobs
616	 */
617	numjobs = td->o.numjobs;
618	while (--numjobs) {
619		struct thread_data *td_new = get_new_job(0, td);
620
621		if (!td_new)
622			goto err;
623
624		td_new->o.numjobs = 1;
625		td_new->o.stonewall = 0;
626		td_new->o.new_group = 0;
627
628		if (file_alloced) {
629			td_new->o.filename = NULL;
630			td_new->files_index = 0;
631			td_new->files_size = 0;
632			td_new->files = NULL;
633		}
634
635		job_add_num = numjobs - 1;
636
637		if (add_job(td_new, jobname, job_add_num))
638			goto err;
639	}
640
641	return 0;
642err:
643	put_job(td);
644	return -1;
645}
646
647/*
648 * Parse as if 'o' was a command line
649 */
650void add_job_opts(const char **o)
651{
652	struct thread_data *td, *td_parent;
653	int i, in_global = 1;
654	char jobname[32];
655
656	i = 0;
657	td_parent = td = NULL;
658	while (o[i]) {
659		if (!strncmp(o[i], "name", 4)) {
660			in_global = 0;
661			if (td)
662				add_job(td, jobname, 0);
663			td = NULL;
664			sprintf(jobname, "%s", o[i] + 5);
665		}
666		if (in_global && !td_parent)
667			td_parent = get_new_job(1, &def_thread);
668		else if (!in_global && !td) {
669			if (!td_parent)
670				td_parent = &def_thread;
671			td = get_new_job(0, td_parent);
672		}
673		if (in_global)
674			fio_options_parse(td_parent, (char **) &o[i], 1);
675		else
676			fio_options_parse(td, (char **) &o[i], 1);
677		i++;
678	}
679
680	if (td)
681		add_job(td, jobname, 0);
682}
683
684static int skip_this_section(const char *name)
685{
686	if (!job_section)
687		return 0;
688	if (!strncmp(name, "global", 6))
689		return 0;
690
691	return strcmp(job_section, name);
692}
693
694static int is_empty_or_comment(char *line)
695{
696	unsigned int i;
697
698	for (i = 0; i < strlen(line); i++) {
699		if (line[i] == ';')
700			return 1;
701		if (line[i] == '#')
702			return 1;
703		if (!isspace(line[i]) && !iscntrl(line[i]))
704			return 0;
705	}
706
707	return 1;
708}
709
710/*
711 * This is our [ini] type file parser.
712 */
713static int parse_jobs_ini(char *file, int stonewall_flag)
714{
715	unsigned int global;
716	struct thread_data *td;
717	char *string, *name;
718	FILE *f;
719	char *p;
720	int ret = 0, stonewall;
721	int first_sect = 1;
722	int skip_fgets = 0;
723	int inside_skip = 0;
724	char **opts;
725	int i, alloc_opts, num_opts;
726
727	if (!strcmp(file, "-"))
728		f = stdin;
729	else
730		f = fopen(file, "r");
731
732	if (!f) {
733		perror("fopen job file");
734		return 1;
735	}
736
737	string = malloc(4096);
738
739	/*
740	 * it's really 256 + small bit, 280 should suffice
741	 */
742	name = malloc(280);
743	memset(name, 0, 280);
744
745	alloc_opts = 8;
746	opts = malloc(sizeof(char *) * alloc_opts);
747	num_opts = 0;
748
749	stonewall = stonewall_flag;
750	do {
751		/*
752		 * if skip_fgets is set, we already have loaded a line we
753		 * haven't handled.
754		 */
755		if (!skip_fgets) {
756			p = fgets(string, 4095, f);
757			if (!p)
758				break;
759		}
760
761		skip_fgets = 0;
762		strip_blank_front(&p);
763		strip_blank_end(p);
764
765		if (is_empty_or_comment(p))
766			continue;
767		if (sscanf(p, "[%255s]", name) != 1) {
768			if (inside_skip)
769				continue;
770			log_err("fio: option <%s> outside of [] job section\n",
771									p);
772			break;
773		}
774
775		if (skip_this_section(name)) {
776			inside_skip = 1;
777			continue;
778		} else
779			inside_skip = 0;
780
781		global = !strncmp(name, "global", 6);
782
783		name[strlen(name) - 1] = '\0';
784
785		if (dump_cmdline) {
786			if (first_sect)
787				log_info("fio ");
788			if (!global)
789				log_info("--name=%s ", name);
790			first_sect = 0;
791		}
792
793		td = get_new_job(global, &def_thread);
794		if (!td) {
795			ret = 1;
796			break;
797		}
798
799		/*
800		 * Seperate multiple job files by a stonewall
801		 */
802		if (!global && stonewall) {
803			td->o.stonewall = stonewall;
804			stonewall = 0;
805		}
806
807		num_opts = 0;
808		memset(opts, 0, alloc_opts * sizeof(char *));
809
810		while ((p = fgets(string, 4096, f)) != NULL) {
811			if (is_empty_or_comment(p))
812				continue;
813
814			strip_blank_front(&p);
815
816			/*
817			 * new section, break out and make sure we don't
818			 * fgets() a new line at the top.
819			 */
820			if (p[0] == '[') {
821				skip_fgets = 1;
822				break;
823			}
824
825			strip_blank_end(p);
826
827			if (num_opts == alloc_opts) {
828				alloc_opts <<= 1;
829				opts = realloc(opts,
830						alloc_opts * sizeof(char *));
831			}
832
833			opts[num_opts] = strdup(p);
834			num_opts++;
835		}
836
837		ret = fio_options_parse(td, opts, num_opts);
838		if (!ret) {
839			if (dump_cmdline)
840				for (i = 0; i < num_opts; i++)
841					log_info("--%s ", opts[i]);
842
843			ret = add_job(td, name, 0);
844		} else {
845			log_err("fio: job %s dropped\n", name);
846			put_job(td);
847		}
848
849		for (i = 0; i < num_opts; i++)
850			free(opts[i]);
851		num_opts = 0;
852	} while (!ret);
853
854	if (dump_cmdline)
855		log_info("\n");
856
857	for (i = 0; i < num_opts; i++)
858		free(opts[i]);
859
860	free(string);
861	free(name);
862	free(opts);
863	if (f != stdin)
864		fclose(f);
865	return ret;
866}
867
868static int fill_def_thread(void)
869{
870	memset(&def_thread, 0, sizeof(def_thread));
871
872	fio_getaffinity(getpid(), &def_thread.o.cpumask);
873
874	/*
875	 * fill default options
876	 */
877	fio_fill_default_options(&def_thread);
878
879	def_thread.o.timeout = def_timeout;
880	def_thread.o.write_bw_log = write_bw_log;
881	def_thread.o.write_lat_log = write_lat_log;
882
883	return 0;
884}
885
886static void free_shm(void)
887{
888	struct shmid_ds sbuf;
889
890	if (threads) {
891		void *tp = threads;
892
893		threads = NULL;
894		file_hash_exit();
895		fio_debug_jobp = NULL;
896		shmdt(tp);
897		shmctl(shm_id, IPC_RMID, &sbuf);
898	}
899
900	scleanup();
901}
902
903/*
904 * The thread area is shared between the main process and the job
905 * threads/processes. So setup a shared memory segment that will hold
906 * all the job info. We use the end of the region for keeping track of
907 * open files across jobs, for file sharing.
908 */
909static int setup_thread_area(void)
910{
911	void *hash;
912
913	/*
914	 * 1024 is too much on some machines, scale max_jobs if
915	 * we get a failure that looks like too large a shm segment
916	 */
917	do {
918		size_t size = max_jobs * sizeof(struct thread_data);
919
920		size += file_hash_size;
921		size += sizeof(unsigned int);
922
923		shm_id = shmget(0, size, IPC_CREAT | 0600);
924		if (shm_id != -1)
925			break;
926		if (errno != EINVAL) {
927			perror("shmget");
928			break;
929		}
930
931		max_jobs >>= 1;
932	} while (max_jobs);
933
934	if (shm_id == -1)
935		return 1;
936
937	threads = shmat(shm_id, NULL, 0);
938	if (threads == (void *) -1) {
939		perror("shmat");
940		return 1;
941	}
942
943	memset(threads, 0, max_jobs * sizeof(struct thread_data));
944	hash = (void *) threads + max_jobs * sizeof(struct thread_data);
945	fio_debug_jobp = (void *) hash + file_hash_size;
946	*fio_debug_jobp = -1;
947	file_hash_init(hash);
948	atexit(free_shm);
949	return 0;
950}
951
952static void usage(const char *name)
953{
954	printf("%s\n", fio_version_string);
955	printf("%s [options] [job options] <job file(s)>\n", name);
956	printf("\t--debug=options\tEnable debug logging\n");
957	printf("\t--output\tWrite output to file\n");
958	printf("\t--timeout\tRuntime in seconds\n");
959	printf("\t--latency-log\tGenerate per-job latency logs\n");
960	printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
961	printf("\t--minimal\tMinimal (terse) output\n");
962	printf("\t--version\tPrint version info and exit\n");
963	printf("\t--help\t\tPrint this page\n");
964	printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
965		" them\n");
966	printf("\t--showcmd\tTurn a job file into command line options\n");
967	printf("\t--eta=when\tWhen ETA estimate should be printed\n");
968	printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
969	printf("\t--readonly\tTurn on safety read-only checks, preventing"
970		" writes\n");
971	printf("\t--section=name\tOnly run specified section in job file\n");
972	printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
973		" (def 1024)\n");
974	printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>\n");
975}
976
977#ifdef FIO_INC_DEBUG
978struct debug_level debug_levels[] = {
979	{ .name = "process",	.shift = FD_PROCESS, },
980	{ .name = "file",	.shift = FD_FILE, },
981	{ .name = "io",		.shift = FD_IO, },
982	{ .name = "mem",	.shift = FD_MEM, },
983	{ .name = "blktrace",	.shift = FD_BLKTRACE },
984	{ .name = "verify",	.shift = FD_VERIFY },
985	{ .name = "random",	.shift = FD_RANDOM },
986	{ .name = "parse",	.shift = FD_PARSE },
987	{ .name = "diskutil",	.shift = FD_DISKUTIL },
988	{ .name = "job",	.shift = FD_JOB },
989	{ .name = "mutex",	.shift = FD_MUTEX },
990	{ .name	= "profile",	.shift = FD_PROFILE },
991	{ .name = NULL, },
992};
993
994static int set_debug(const char *string)
995{
996	struct debug_level *dl;
997	char *p = (char *) string;
998	char *opt;
999	int i;
1000
1001	if (!strcmp(string, "?") || !strcmp(string, "help")) {
1002		int i;
1003
1004		log_info("fio: dumping debug options:");
1005		for (i = 0; debug_levels[i].name; i++) {
1006			dl = &debug_levels[i];
1007			log_info("%s,", dl->name);
1008		}
1009		log_info("all\n");
1010		return 1;
1011	}
1012
1013	while ((opt = strsep(&p, ",")) != NULL) {
1014		int found = 0;
1015
1016		if (!strncmp(opt, "all", 3)) {
1017			log_info("fio: set all debug options\n");
1018			fio_debug = ~0UL;
1019			continue;
1020		}
1021
1022		for (i = 0; debug_levels[i].name; i++) {
1023			dl = &debug_levels[i];
1024			found = !strncmp(opt, dl->name, strlen(dl->name));
1025			if (!found)
1026				continue;
1027
1028			if (dl->shift == FD_JOB) {
1029				opt = strchr(opt, ':');
1030				if (!opt) {
1031					log_err("fio: missing job number\n");
1032					break;
1033				}
1034				opt++;
1035				fio_debug_jobno = atoi(opt);
1036				log_info("fio: set debug jobno %d\n",
1037							fio_debug_jobno);
1038			} else {
1039				log_info("fio: set debug option %s\n", opt);
1040				fio_debug |= (1UL << dl->shift);
1041			}
1042			break;
1043		}
1044
1045		if (!found)
1046			log_err("fio: debug mask %s not found\n", opt);
1047	}
1048	return 0;
1049}
1050#else
1051static int set_debug(const char *string)
1052{
1053	log_err("fio: debug tracing not included in build\n");
1054	return 1;
1055}
1056#endif
1057
1058static int parse_cmd_line(int argc, char *argv[])
1059{
1060	struct thread_data *td = NULL;
1061	int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1062
1063	while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1064		switch (c) {
1065		case 'a':
1066			smalloc_pool_size = atoi(optarg);
1067			break;
1068		case 't':
1069			def_timeout = atoi(optarg);
1070			break;
1071		case 'l':
1072			write_lat_log = 1;
1073			break;
1074		case 'w':
1075			write_bw_log = 1;
1076			break;
1077		case 'o':
1078			f_out = fopen(optarg, "w+");
1079			if (!f_out) {
1080				perror("fopen output");
1081				exit(1);
1082			}
1083			f_err = f_out;
1084			break;
1085		case 'm':
1086			terse_output = 1;
1087			break;
1088		case 'h':
1089			usage(argv[0]);
1090			exit(0);
1091		case 'c':
1092			exit(fio_show_option_help(optarg));
1093		case 's':
1094			dump_cmdline = 1;
1095			break;
1096		case 'r':
1097			read_only = 1;
1098			break;
1099		case 'v':
1100			printf("%s\n", fio_version_string);
1101			exit(0);
1102		case 'e':
1103			if (!strcmp("always", optarg))
1104				eta_print = FIO_ETA_ALWAYS;
1105			else if (!strcmp("never", optarg))
1106				eta_print = FIO_ETA_NEVER;
1107			break;
1108		case 'd':
1109			if (set_debug(optarg))
1110				do_exit++;
1111			break;
1112		case 'x':
1113			if (!strcmp(optarg, "global")) {
1114				log_err("fio: can't use global as only "
1115					"section\n");
1116				do_exit++;
1117				exit_val = 1;
1118				break;
1119			}
1120			if (job_section)
1121				free(job_section);
1122			job_section = strdup(optarg);
1123			break;
1124		case 'p':
1125			exec_profile = strdup(optarg);
1126			break;
1127		case FIO_GETOPT_JOB: {
1128			const char *opt = l_opts[lidx].name;
1129			char *val = optarg;
1130
1131			if (!strncmp(opt, "name", 4) && td) {
1132				ret = add_job(td, td->o.name ?: "fio", 0);
1133				if (ret) {
1134					put_job(td);
1135					return 0;
1136				}
1137				td = NULL;
1138			}
1139			if (!td) {
1140				int is_section = !strncmp(opt, "name", 4);
1141				int global = 0;
1142
1143				if (!is_section || !strncmp(val, "global", 6))
1144					global = 1;
1145
1146				if (is_section && skip_this_section(val))
1147					continue;
1148
1149				td = get_new_job(global, &def_thread);
1150				if (!td)
1151					return 0;
1152			}
1153
1154			ret = fio_cmd_option_parse(td, opt, val);
1155			break;
1156		}
1157		default:
1158			do_exit++;
1159			exit_val = 1;
1160			break;
1161		}
1162	}
1163
1164	if (do_exit)
1165		exit(exit_val);
1166
1167	if (td) {
1168		if (!ret)
1169			ret = add_job(td, td->o.name ?: "fio", 0);
1170		if (ret)
1171			put_job(td);
1172	}
1173
1174	while (optind < argc) {
1175		ini_idx++;
1176		ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1177		ini_file[ini_idx - 1] = strdup(argv[optind]);
1178		optind++;
1179	}
1180
1181	return ini_idx;
1182}
1183
1184int parse_options(int argc, char *argv[])
1185{
1186	int job_files, i;
1187
1188	f_out = stdout;
1189	f_err = stderr;
1190
1191	fio_options_dup_and_init(l_opts);
1192
1193	if (setup_thread_area())
1194		return 1;
1195	if (fill_def_thread())
1196		return 1;
1197
1198	job_files = parse_cmd_line(argc, argv);
1199
1200	for (i = 0; i < job_files; i++) {
1201		if (fill_def_thread())
1202			return 1;
1203		if (parse_jobs_ini(ini_file[i], i))
1204			return 1;
1205		free(ini_file[i]);
1206	}
1207
1208	free(ini_file);
1209	options_mem_free(&def_thread);
1210
1211	if (!thread_number) {
1212		if (dump_cmdline)
1213			return 0;
1214		if (exec_profile)
1215			return 0;
1216
1217		log_err("No jobs defined(s)\n\n");
1218		usage(argv[0]);
1219		return 1;
1220	}
1221
1222	if (def_thread.o.gtod_offload) {
1223		fio_gtod_init();
1224		fio_gtod_offload = 1;
1225		fio_gtod_cpu = def_thread.o.gtod_cpu;
1226	}
1227
1228	return 0;
1229}
1230