init.c revision 245142ff7554290cf666e3ef6b7b57512bf17d75 
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 <assert.h>
13#include <sys/ipc.h>
14#include <sys/shm.h>
15#include <sys/types.h>
16#include <sys/stat.h>
17
18#include "fio.h"
19#include "parse.h"
20
21/*
22 * The default options
23 */
24#define DEF_BS			(4096)
25#define DEF_TIMEOUT		(0)
26#define DEF_RATE_CYCLE		(1000)
27#define DEF_ODIRECT		(1)
28#define DEF_IO_ENGINE		(FIO_SYNCIO)
29#define DEF_IO_ENGINE_NAME	"sync"
30#define DEF_SEQUENTIAL		(1)
31#define DEF_RAND_REPEAT		(1)
32#define DEF_OVERWRITE		(1)
33#define DEF_INVALIDATE		(1)
34#define DEF_SYNCIO		(0)
35#define DEF_RANDSEED		(0xb1899bedUL)
36#define DEF_BWAVGTIME		(500)
37#define DEF_CREATE_SER		(1)
38#define DEF_CREATE_FSYNC	(1)
39#define DEF_LOOPS		(1)
40#define DEF_VERIFY		(0)
41#define DEF_STONEWALL		(0)
42#define DEF_NUMJOBS		(1)
43#define DEF_USE_THREAD		(0)
44#define DEF_FILE_SIZE		(1024 * 1024 * 1024UL)
45#define DEF_ZONE_SIZE		(0)
46#define DEF_ZONE_SKIP		(0)
47#define DEF_RWMIX_CYCLE		(500)
48#define DEF_RWMIX_READ		(50)
49#define DEF_NICE		(0)
50#define DEF_NR_FILES		(1)
51#define DEF_UNLINK		(0)
52#define DEF_WRITE_BW_LOG	(0)
53#define DEF_WRITE_LAT_LOG	(0)
54#define DEF_NO_RAND_MAP		(0)
55
56#define td_var_offset(var)	((size_t) &((struct thread_data *)0)->var)
57
58static int str_rw_cb(void *, const char *);
59static int str_ioengine_cb(void *, const char *);
60static int str_mem_cb(void *, const char *);
61static int str_verify_cb(void *, const char *);
62static int str_lockmem_cb(void *, unsigned long *);
63#ifdef FIO_HAVE_IOPRIO
64static int str_prio_cb(void *, unsigned int *);
65static int str_prioclass_cb(void *, unsigned int *);
66#endif
67static int str_exitall_cb(void);
68static int str_cpumask_cb(void *, unsigned int *);
69
70/*
71 * Map of job/command line options
72 */
73static struct fio_option options[] = {
74	{
75		.name	= "name",
76		.type	= FIO_OPT_STR_STORE,
77		.off1	= td_var_offset(name),
78	},
79	{
80		.name	= "directory",
81		.type	= FIO_OPT_STR_STORE,
82		.off1	= td_var_offset(directory),
83	},
84	{
85		.name	= "filename",
86		.type	= FIO_OPT_STR_STORE,
87		.off1	= td_var_offset(filename),
88	},
89	{
90		.name	= "rw",
91		.type	= FIO_OPT_STR,
92		.cb	= str_rw_cb,
93	},
94	{
95		.name	= "ioengine",
96		.type	= FIO_OPT_STR,
97		.cb	= str_ioengine_cb,
98	},
99	{
100		.name	= "mem",
101		.type	= FIO_OPT_STR,
102		.cb	= str_mem_cb,
103	},
104	{
105		.name	= "verify",
106		.type	= FIO_OPT_STR,
107		.cb	= str_verify_cb,
108	},
109	{
110		.name	= "write_iolog",
111		.type	= FIO_OPT_STR_STORE,
112		.off1	= td_var_offset(write_iolog_file),
113	},
114	{
115		.name	= "read_iolog",
116		.type	= FIO_OPT_STR_STORE,
117		.off1	= td_var_offset(read_iolog_file),
118	},
119	{
120		.name	= "exec_prerun",
121		.type	= FIO_OPT_STR_STORE,
122		.off1	= td_var_offset(exec_prerun),
123	},
124	{
125		.name	= "exec_postrun",
126		.type	= FIO_OPT_STR_STORE,
127		.off1	= td_var_offset(exec_postrun),
128	},
129#ifdef FIO_HAVE_IOSCHED_SWITCH
130	{
131		.name	= "ioscheduler",
132		.type	= FIO_OPT_STR_STORE,
133		.off1	= td_var_offset(ioscheduler),
134	},
135#endif
136	{
137		.name	= "size",
138		.type	= FIO_OPT_STR_VAL,
139		.off1	= td_var_offset(total_file_size),
140	},
141	{
142		.name	= "bs",
143		.type	= FIO_OPT_STR_VAL_INT,
144		.off1	= td_var_offset(bs[DDIR_READ]),
145		.off2	= td_var_offset(bs[DDIR_WRITE]),
146	},
147	{
148		.name	= "offset",
149		.type	= FIO_OPT_STR_VAL,
150		.off1	= td_var_offset(start_offset),
151	},
152	{
153		.name	= "zonesize",
154		.type	= FIO_OPT_STR_VAL,
155		.off1	= td_var_offset(zone_size),
156	},
157	{
158		.name	= "zoneskip",
159		.type	= FIO_OPT_STR_VAL,
160		.off1	= td_var_offset(zone_skip),
161	},
162	{
163		.name	= "lockmem",
164		.type	= FIO_OPT_STR_VAL,
165		.cb	= str_lockmem_cb,
166	},
167	{
168		.name	= "bsrange",
169		.type	= FIO_OPT_RANGE,
170		.off1	= td_var_offset(min_bs[DDIR_READ]),
171		.off2	= td_var_offset(max_bs[DDIR_READ]),
172		.off3	= td_var_offset(min_bs[DDIR_WRITE]),
173		.off4	= td_var_offset(max_bs[DDIR_WRITE]),
174	},
175	{
176		.name	= "nrfiles",
177		.type	= FIO_OPT_INT,
178		.off1	= td_var_offset(nr_files),
179	},
180	{
181		.name	= "iodepth",
182		.type	= FIO_OPT_INT,
183		.off1	= td_var_offset(iodepth),
184	},
185	{
186		.name	= "fsync",
187		.type	= FIO_OPT_INT,
188		.off1	= td_var_offset(fsync_blocks),
189	},
190	{
191		.name	= "rwmixcycle",
192		.type	= FIO_OPT_INT,
193		.off1	= td_var_offset(rwmixcycle),
194	},
195	{
196		.name	= "rwmixread",
197		.type	= FIO_OPT_INT,
198		.off1	= td_var_offset(rwmixread),
199		.max_val= 100,
200	},
201	{
202		.name	= "rwmixwrite",
203		.type	= FIO_OPT_INT,
204		.off1	= td_var_offset(rwmixwrite),
205		.max_val= 100,
206	},
207	{
208		.name	= "nice",
209		.type	= FIO_OPT_INT,
210		.off1	= td_var_offset(nice),
211	},
212#ifdef FIO_HAVE_IOPRIO
213	{
214		.name	= "prio",
215		.type	= FIO_OPT_INT,
216		.cb	= str_prio_cb,
217	},
218	{
219		.name	= "prioclass",
220		.type	= FIO_OPT_INT,
221		.cb	= str_prioclass_cb,
222	},
223#endif
224	{
225		.name	= "thinktime",
226		.type	= FIO_OPT_INT,
227		.off1	= td_var_offset(thinktime)
228	},
229	{
230		.name	= "rate",
231		.type	= FIO_OPT_INT,
232		.off1	= td_var_offset(rate)
233	},
234	{
235		.name	= "ratemin",
236		.type	= FIO_OPT_INT,
237		.off1	= td_var_offset(ratemin)
238	},
239	{
240		.name	= "ratecycle",
241		.type	= FIO_OPT_INT,
242		.off1	= td_var_offset(ratecycle)
243	},
244	{
245		.name	= "startdelay",
246		.type	= FIO_OPT_INT,
247		.off1	= td_var_offset(start_delay)
248	},
249	{
250		.name	= "timeout",
251		.type	= FIO_OPT_STR_VAL_TIME,
252		.off1	= td_var_offset(timeout)
253	},
254	{
255		.name	= "invalidate",
256		.type	= FIO_OPT_INT,
257		.off1	= td_var_offset(invalidate_cache)
258	},
259	{
260		.name	= "sync",
261		.type	= FIO_OPT_INT,
262		.off1	= td_var_offset(sync_io)
263	},
264	{
265		.name	= "bwavgtime",
266		.type	= FIO_OPT_INT,
267		.off1	= td_var_offset(bw_avg_time)
268	},
269	{
270		.name	= "create_serialize",
271		.type	= FIO_OPT_INT,
272		.off1	= td_var_offset(create_serialize)
273	},
274	{
275		.name	= "create_fsync",
276		.type	= FIO_OPT_INT,
277		.off1	= td_var_offset(create_fsync)
278	},
279	{
280		.name	= "loops",
281		.type	= FIO_OPT_INT,
282		.off1	= td_var_offset(loops)
283	},
284	{
285		.name	= "numjobs",
286		.type	= FIO_OPT_INT,
287		.off1	= td_var_offset(numjobs)
288	},
289	{
290		.name	= "cpuload",
291		.type	= FIO_OPT_INT,
292		.off1	= td_var_offset(cpuload)
293	},
294	{
295		.name	= "cpuchunks",
296		.type	= FIO_OPT_INT,
297		.off1	= td_var_offset(cpucycle)
298	},
299	{
300		.name	= "direct",
301		.type	= FIO_OPT_INT,
302		.off1	= td_var_offset(odirect)
303	},
304	{
305		.name	= "overwrite",
306		.type	= FIO_OPT_INT,
307		.off1	= td_var_offset(overwrite)
308	},
309#ifdef FIO_HAVE_CPU_AFFINITY
310	{
311		.name	= "cpumask",
312		.type	= FIO_OPT_INT,
313		.cb	= str_cpumask_cb,
314	},
315#endif
316	{
317		.name	= "end_fsync",
318		.type	= FIO_OPT_INT,
319		.off1	= td_var_offset(end_fsync)
320	},
321	{
322		.name	= "unlink",
323		.type	= FIO_OPT_STR_SET,
324		.off1	= td_var_offset(unlink),
325	},
326	{
327		.name	= "exitall",
328		.type	= FIO_OPT_STR_SET,
329		.cb	= str_exitall_cb,
330	},
331	{
332		.name	= "stonewall",
333		.type	= FIO_OPT_STR_SET,
334		.off1	= td_var_offset(stonewall),
335	},
336	{
337		.name	= "thread",
338		.type	= FIO_OPT_STR_SET,
339		.off1	= td_var_offset(thread),
340	},
341	{
342		.name	= "write_bw_log",
343		.type	= FIO_OPT_STR_SET,
344		.off1	= td_var_offset(write_bw_log),
345	},
346	{
347		.name	= "write_lat_log",
348		.type	= FIO_OPT_STR_SET,
349		.off1	= td_var_offset(write_lat_log),
350	},
351	{
352		.name	= "norandommap",
353		.type	= FIO_OPT_STR_SET,
354		.off1	= td_var_offset(norandommap),
355	},
356	{
357		.name	= "bs_unaligned",
358		.type	= FIO_OPT_STR_SET,
359		.off1	= td_var_offset(bs_unaligned),
360	},
361	{
362		.name = NULL,
363	},
364};
365
366#define FIO_JOB_OPTS	(sizeof(options) / sizeof(struct fio_option))
367#define FIO_CMD_OPTS	(16)
368#define FIO_GETOPT_JOB	(0x89988998)
369
370/*
371 * Command line options. These will contain the above, plus a few
372 * extra that only pertain to fio itself and not jobs.
373 */
374static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
375	{
376		.name		= "output",
377		.has_arg	= required_argument,
378		.val		= 'o',
379	},
380	{
381		.name		= "timeout",
382		.has_arg	= required_argument,
383		.val		= 't',
384	},
385	{
386		.name		= "latency-log",
387		.has_arg	= required_argument,
388		.val		= 'l',
389	},
390	{
391		.name		= "bandwidth-log",
392		.has_arg	= required_argument,
393		.val		= 'b',
394	},
395	{
396		.name		= "minimal",
397		.has_arg	= optional_argument,
398		.val		= 'm',
399	},
400	{
401		.name		= "version",
402		.has_arg	= no_argument,
403		.val		= 'v',
404	},
405	{
406		.name		= NULL,
407	},
408};
409
410static int def_timeout = DEF_TIMEOUT;
411
412static char fio_version_string[] = "fio 1.9";
413
414static char **ini_file;
415static int max_jobs = MAX_JOBS;
416
417struct thread_data def_thread;
418struct thread_data *threads = NULL;
419
420int exitall_on_terminate = 0;
421int terse_output = 0;
422unsigned long long mlock_size = 0;
423FILE *f_out = NULL;
424FILE *f_err = NULL;
425
426static int write_lat_log = DEF_WRITE_LAT_LOG;
427static int write_bw_log = DEF_WRITE_BW_LOG;
428
429/*
430 * Return a free job structure.
431 */
432static struct thread_data *get_new_job(int global, struct thread_data *parent)
433{
434	struct thread_data *td;
435
436	if (global)
437		return &def_thread;
438	if (thread_number >= max_jobs)
439		return NULL;
440
441	td = &threads[thread_number++];
442	*td = *parent;
443
444	td->thread_number = thread_number;
445	return td;
446}
447
448static void put_job(struct thread_data *td)
449{
450	if (td == &def_thread)
451		return;
452
453	memset(&threads[td->thread_number - 1], 0, sizeof(*td));
454	thread_number--;
455}
456
457/*
458 * Lazy way of fixing up options that depend on each other. We could also
459 * define option callback handlers, but this is easier.
460 */
461static void fixup_options(struct thread_data *td)
462{
463	if (!td->rwmixread && td->rwmixwrite)
464		td->rwmixread = 100 - td->rwmixwrite;
465
466	if (td->write_iolog_file && td->read_iolog_file) {
467		log_err("fio: read iolog overrides write_iolog\n");
468		free(td->write_iolog_file);
469		td->write_iolog_file = NULL;
470	}
471
472	if (td->io_ops->flags & FIO_SYNCIO)
473		td->iodepth = 1;
474	else {
475		if (!td->iodepth)
476			td->iodepth = td->nr_files;
477	}
478
479	/*
480	 * only really works for sequential io for now, and with 1 file
481	 */
482	if (td->zone_size && !td->sequential && td->nr_files == 1)
483		td->zone_size = 0;
484
485	/*
486	 * Reads can do overwrites, we always need to pre-create the file
487	 */
488	if (td_read(td) || td_rw(td))
489		td->overwrite = 1;
490
491	if (!td->min_bs[DDIR_READ])
492		td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
493	if (!td->max_bs[DDIR_READ])
494		td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
495	if (!td->min_bs[DDIR_WRITE])
496		td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
497	if (!td->max_bs[DDIR_WRITE])
498		td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
499
500	td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
501
502	if (td_read(td) && !td_rw(td))
503		td->verify = 0;
504
505	if (td->norandommap && td->verify != VERIFY_NONE) {
506		log_err("fio: norandommap given, verify disabled\n");
507		td->verify = VERIFY_NONE;
508	}
509	if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
510		log_err("fio: bs_unaligned may not work with raw io\n");
511}
512
513/*
514 * This function leaks the buffer
515 */
516static char *to_kmg(unsigned int val)
517{
518	char *buf = malloc(32);
519	char post[] = { 0, 'K', 'M', 'G', 'P', 0 };
520	char *p = post;
521
522	do {
523		if (val & 1023)
524			break;
525
526		val >>= 10;
527		p++;
528	} while (*p);
529
530	snprintf(buf, 31, "%u%c", val, *p);
531	return buf;
532}
533
534/*
535 * Adds a job to the list of things todo. Sanitizes the various options
536 * to make sure we don't have conflicts, and initializes various
537 * members of td.
538 */
539static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
540{
541	const char *ddir_str[] = { "read", "write", "randread", "randwrite",
542				   "rw", NULL, "randrw" };
543	struct stat sb;
544	int numjobs, ddir, i;
545	struct fio_file *f;
546
547	/*
548	 * the def_thread is just for options, it's not a real job
549	 */
550	if (td == &def_thread)
551		return 0;
552
553	/*
554	 * Set default io engine, if none set
555	 */
556	if (!td->io_ops) {
557		td->io_ops = load_ioengine(td, DEF_IO_ENGINE_NAME);
558		if (!td->io_ops) {
559			log_err("default engine %s not there?\n", DEF_IO_ENGINE_NAME);
560			return 1;
561		}
562	}
563
564	if (td->odirect)
565		td->io_ops->flags |= FIO_RAWIO;
566
567	fixup_options(td);
568
569	td->filetype = FIO_TYPE_FILE;
570	if (!stat(jobname, &sb)) {
571		if (S_ISBLK(sb.st_mode))
572			td->filetype = FIO_TYPE_BD;
573		else if (S_ISCHR(sb.st_mode))
574			td->filetype = FIO_TYPE_CHAR;
575	}
576
577	if (td->filename)
578		td->nr_uniq_files = 1;
579	else
580		td->nr_uniq_files = td->nr_files;
581
582	if (td->filetype == FIO_TYPE_FILE || td->filename) {
583		char tmp[PATH_MAX];
584		int len = 0;
585		int i;
586
587		if (td->directory && td->directory[0] != '\0')
588			sprintf(tmp, "%s/", td->directory);
589
590		td->files = malloc(sizeof(struct fio_file) * td->nr_files);
591
592		for_each_file(td, f, i) {
593			memset(f, 0, sizeof(*f));
594			f->fd = -1;
595
596			if (td->filename)
597				sprintf(tmp + len, "%s", td->filename);
598			else
599				sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
600			f->file_name = strdup(tmp);
601		}
602	} else {
603		td->nr_files = 1;
604		td->files = malloc(sizeof(struct fio_file));
605		f = &td->files[0];
606
607		memset(f, 0, sizeof(*f));
608		f->fd = -1;
609		f->file_name = strdup(jobname);
610	}
611
612	for_each_file(td, f, i) {
613		f->file_size = td->total_file_size / td->nr_files;
614		f->file_offset = td->start_offset;
615	}
616
617	fio_sem_init(&td->mutex, 0);
618
619	td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
620	td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
621	td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
622
623	if (td->stonewall && td->thread_number > 1)
624		groupid++;
625
626	td->groupid = groupid;
627
628	if (setup_rate(td))
629		goto err;
630
631	if (td->write_lat_log) {
632		setup_log(&td->slat_log);
633		setup_log(&td->clat_log);
634	}
635	if (td->write_bw_log)
636		setup_log(&td->bw_log);
637
638	if (!td->name)
639		td->name = strdup(jobname);
640
641	ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
642
643	if (!terse_output) {
644		if (!job_add_num) {
645			if (td->io_ops->flags & FIO_CPUIO)
646				fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
647			else {
648				char *c1, *c2, *c3, *c4;
649
650				c1 = to_kmg(td->min_bs[DDIR_READ]);
651				c2 = to_kmg(td->max_bs[DDIR_READ]);
652				c3 = to_kmg(td->min_bs[DDIR_WRITE]);
653				c4 = to_kmg(td->max_bs[DDIR_WRITE]);
654
655				fprintf(f_out, "%s: (g=%d): rw=%s, odir=%d, bs=%s-%s/%s-%s, rate=%d, ioengine=%s, iodepth=%d\n", td->name, td->groupid, ddir_str[ddir], td->odirect, c1, c2, c3, c4, td->rate, td->io_ops->name, td->iodepth);
656
657				free(c1);
658				free(c2);
659				free(c3);
660				free(c4);
661			}
662		} else if (job_add_num == 1)
663			fprintf(f_out, "...\n");
664	}
665
666	/*
667	 * recurse add identical jobs, clear numjobs and stonewall options
668	 * as they don't apply to sub-jobs
669	 */
670	numjobs = td->numjobs;
671	while (--numjobs) {
672		struct thread_data *td_new = get_new_job(0, td);
673
674		if (!td_new)
675			goto err;
676
677		td_new->numjobs = 1;
678		td_new->stonewall = 0;
679		job_add_num = numjobs - 1;
680
681		if (add_job(td_new, jobname, job_add_num))
682			goto err;
683	}
684	return 0;
685err:
686	put_job(td);
687	return -1;
688}
689
690/*
691 * Initialize the various random states we need (random io, block size ranges,
692 * read/write mix, etc).
693 */
694int init_random_state(struct thread_data *td)
695{
696	unsigned long seeds[4];
697	int fd, num_maps, blocks, i;
698	struct fio_file *f;
699
700	if (td->io_ops->flags & FIO_CPUIO)
701		return 0;
702
703	fd = open("/dev/urandom", O_RDONLY);
704	if (fd == -1) {
705		td_verror(td, errno);
706		return 1;
707	}
708
709	if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
710		td_verror(td, EIO);
711		close(fd);
712		return 1;
713	}
714
715	close(fd);
716
717	os_random_seed(seeds[0], &td->bsrange_state);
718	os_random_seed(seeds[1], &td->verify_state);
719	os_random_seed(seeds[2], &td->rwmix_state);
720
721	if (td->sequential)
722		return 0;
723
724	if (td->rand_repeatable)
725		seeds[3] = DEF_RANDSEED;
726
727	if (!td->norandommap) {
728		for_each_file(td, f, i) {
729			blocks = (f->file_size + td->rw_min_bs - 1) / td->rw_min_bs;
730			num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
731			f->file_map = malloc(num_maps * sizeof(long));
732			f->num_maps = num_maps;
733			memset(f->file_map, 0, num_maps * sizeof(long));
734		}
735	}
736
737	os_random_seed(seeds[3], &td->random_state);
738	return 0;
739}
740
741static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
742{
743#ifdef FIO_HAVE_CPU_AFFINITY
744	unsigned int i;
745
746	CPU_ZERO(&cpumask);
747
748	for (i = 0; i < sizeof(int) * 8; i++) {
749		if ((1 << i) & cpu)
750			CPU_SET(i, &cpumask);
751	}
752#endif
753}
754
755static int is_empty_or_comment(char *line)
756{
757	unsigned int i;
758
759	for (i = 0; i < strlen(line); i++) {
760		if (line[i] == ';')
761			return 1;
762		if (!isspace(line[i]) && !iscntrl(line[i]))
763			return 0;
764	}
765
766	return 1;
767}
768
769static int str_rw_cb(void *data, const char *mem)
770{
771	struct thread_data *td = data;
772
773	if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
774		td->ddir = DDIR_READ;
775		td->sequential = 1;
776		return 0;
777	} else if (!strncmp(mem, "randread", 8)) {
778		td->ddir = DDIR_READ;
779		td->sequential = 0;
780		return 0;
781	} else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
782		td->ddir = DDIR_WRITE;
783		td->sequential = 1;
784		return 0;
785	} else if (!strncmp(mem, "randwrite", 9)) {
786		td->ddir = DDIR_WRITE;
787		td->sequential = 0;
788		return 0;
789	} else if (!strncmp(mem, "rw", 2)) {
790		td->ddir = 0;
791		td->iomix = 1;
792		td->sequential = 1;
793		return 0;
794	} else if (!strncmp(mem, "randrw", 6)) {
795		td->ddir = 0;
796		td->iomix = 1;
797		td->sequential = 0;
798		return 0;
799	}
800
801	log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
802	return 1;
803}
804
805static int str_verify_cb(void *data, const char *mem)
806{
807	struct thread_data *td = data;
808
809	if (!strncmp(mem, "0", 1)) {
810		td->verify = VERIFY_NONE;
811		return 0;
812	} else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
813		td->verify = VERIFY_MD5;
814		return 0;
815	} else if (!strncmp(mem, "crc32", 5)) {
816		td->verify = VERIFY_CRC32;
817		return 0;
818	}
819
820	log_err("fio: verify types: md5, crc32\n");
821	return 1;
822}
823
824static int str_mem_cb(void *data, const char *mem)
825{
826	struct thread_data *td = data;
827
828	if (!strncmp(mem, "malloc", 6)) {
829		td->mem_type = MEM_MALLOC;
830		return 0;
831	} else if (!strncmp(mem, "shm", 3)) {
832		td->mem_type = MEM_SHM;
833		return 0;
834	} else if (!strncmp(mem, "mmap", 4)) {
835		td->mem_type = MEM_MMAP;
836		return 0;
837	}
838
839	log_err("fio: mem type: malloc, shm, mmap\n");
840	return 1;
841}
842
843static int str_ioengine_cb(void *data, const char *str)
844{
845	struct thread_data *td = data;
846
847	td->io_ops = load_ioengine(td, str);
848	if (td->io_ops)
849		return 0;
850
851	log_err("fio: ioengine: { linuxaio, aio, libaio }, posixaio, sync, mmap, sgio, splice, cpu\n");
852	log_err("fio: or specify path to dynamic ioengine module\n");
853	return 1;
854}
855
856static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
857{
858	mlock_size = *val;
859	return 0;
860}
861
862#ifdef FIO_HAVE_IOPRIO
863static int str_prioclass_cb(void *data, unsigned int *val)
864{
865	struct thread_data *td = data;
866
867	td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
868	return 0;
869}
870
871static int str_prio_cb(void *data, unsigned int *val)
872{
873	struct thread_data *td = data;
874
875	td->ioprio |= *val;
876	return 0;
877}
878#endif
879
880static int str_exitall_cb(void)
881{
882	exitall_on_terminate = 1;
883	return 0;
884}
885
886static int str_cpumask_cb(void *data, unsigned int *val)
887{
888	struct thread_data *td = data;
889
890	fill_cpu_mask(td->cpumask, *val);
891	return 0;
892}
893
894/*
895 * This is our [ini] type file parser.
896 */
897int parse_jobs_ini(char *file, int stonewall_flag)
898{
899	unsigned int global;
900	struct thread_data *td;
901	char *string, *name;
902	fpos_t off;
903	FILE *f;
904	char *p;
905	int ret = 0, stonewall;
906
907	f = fopen(file, "r");
908	if (!f) {
909		perror("fopen job file");
910		return 1;
911	}
912
913	string = malloc(4096);
914	name = malloc(256);
915	memset(name, 0, 256);
916
917	stonewall = stonewall_flag;
918	do {
919		p = fgets(string, 4095, f);
920		if (!p)
921			break;
922		if (is_empty_or_comment(p))
923			continue;
924		if (sscanf(p, "[%255s]", name) != 1)
925			continue;
926
927		global = !strncmp(name, "global", 6);
928
929		name[strlen(name) - 1] = '\0';
930
931		td = get_new_job(global, &def_thread);
932		if (!td) {
933			ret = 1;
934			break;
935		}
936
937		/*
938		 * Seperate multiple job files by a stonewall
939		 */
940		if (!global && stonewall) {
941			td->stonewall = stonewall;
942			stonewall = 0;
943		}
944
945		fgetpos(f, &off);
946		while ((p = fgets(string, 4096, f)) != NULL) {
947			if (is_empty_or_comment(p))
948				continue;
949
950			strip_blank_front(&p);
951
952			if (p[0] == '[')
953				break;
954
955			strip_blank_end(p);
956
957			fgetpos(f, &off);
958
959			/*
960			 * Don't break here, continue parsing options so we
961			 * dump all the bad ones. Makes trial/error fixups
962			 * easier on the user.
963			 */
964			ret |= parse_option(p, options, td);
965		}
966
967		if (!ret) {
968			fsetpos(f, &off);
969			ret = add_job(td, name, 0);
970		} else {
971			log_err("fio: job %s dropped\n", name);
972			put_job(td);
973		}
974	} while (!ret);
975
976	free(string);
977	free(name);
978	fclose(f);
979	return ret;
980}
981
982static int fill_def_thread(void)
983{
984	memset(&def_thread, 0, sizeof(def_thread));
985
986	if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
987		perror("sched_getaffinity");
988		return 1;
989	}
990
991	/*
992	 * fill globals
993	 */
994	def_thread.ddir = DDIR_READ;
995	def_thread.iomix = 0;
996	def_thread.bs[DDIR_READ] = DEF_BS;
997	def_thread.bs[DDIR_WRITE] = DEF_BS;
998	def_thread.min_bs[DDIR_READ] = def_thread.min_bs[DDIR_WRITE] = 0;
999	def_thread.max_bs[DDIR_READ] = def_thread.max_bs[DDIR_WRITE] = 0;
1000	def_thread.odirect = DEF_ODIRECT;
1001	def_thread.ratecycle = DEF_RATE_CYCLE;
1002	def_thread.sequential = DEF_SEQUENTIAL;
1003	def_thread.timeout = def_timeout;
1004	def_thread.overwrite = DEF_OVERWRITE;
1005	def_thread.invalidate_cache = DEF_INVALIDATE;
1006	def_thread.sync_io = DEF_SYNCIO;
1007	def_thread.mem_type = MEM_MALLOC;
1008	def_thread.bw_avg_time = DEF_BWAVGTIME;
1009	def_thread.create_serialize = DEF_CREATE_SER;
1010	def_thread.create_fsync = DEF_CREATE_FSYNC;
1011	def_thread.loops = DEF_LOOPS;
1012	def_thread.verify = DEF_VERIFY;
1013	def_thread.stonewall = DEF_STONEWALL;
1014	def_thread.numjobs = DEF_NUMJOBS;
1015	def_thread.use_thread = DEF_USE_THREAD;
1016	def_thread.rwmixcycle = DEF_RWMIX_CYCLE;
1017	def_thread.rwmixread = DEF_RWMIX_READ;
1018	def_thread.nice = DEF_NICE;
1019	def_thread.rand_repeatable = DEF_RAND_REPEAT;
1020	def_thread.nr_files = DEF_NR_FILES;
1021	def_thread.unlink = DEF_UNLINK;
1022	def_thread.write_bw_log = write_bw_log;
1023	def_thread.write_lat_log = write_lat_log;
1024	def_thread.norandommap = DEF_NO_RAND_MAP;
1025#ifdef FIO_HAVE_DISK_UTIL
1026	def_thread.do_disk_util = 1;
1027#endif
1028
1029	return 0;
1030}
1031
1032static void usage(void)
1033{
1034	printf("%s\n", fio_version_string);
1035	printf("\t--output\tWrite output to file\n");
1036	printf("\t--timeout\tRuntime in seconds\n");
1037	printf("\t--latency-log\tGenerate per-job latency logs\n");
1038	printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1039	printf("\t--minimal\tMinimal (terse) output\n");
1040	printf("\t--version\tPrint version info and exit\n");
1041}
1042
1043static int parse_cmd_line(int argc, char *argv[])
1044{
1045	struct thread_data *td = NULL;
1046	int c, ini_idx = 0, lidx, ret;
1047
1048	while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1049		switch (c) {
1050		case 't':
1051			def_timeout = atoi(optarg);
1052			break;
1053		case 'l':
1054			write_lat_log = 1;
1055			break;
1056		case 'w':
1057			write_bw_log = 1;
1058			break;
1059		case 'o':
1060			f_out = fopen(optarg, "w+");
1061			if (!f_out) {
1062				perror("fopen output");
1063				exit(1);
1064			}
1065			f_err = f_out;
1066			break;
1067		case 'm':
1068			terse_output = 1;
1069			break;
1070		case 'h':
1071			usage();
1072			exit(0);
1073		case 'v':
1074			printf("%s\n", fio_version_string);
1075			exit(0);
1076		case FIO_GETOPT_JOB: {
1077			const char *opt = long_options[lidx].name;
1078			char *val = optarg;
1079
1080			if (!strncmp(opt, "name", 4) && td) {
1081				ret = add_job(td, td->name ?: "fio", 0);
1082				if (ret) {
1083					put_job(td);
1084					return 0;
1085				}
1086				td = NULL;
1087			}
1088			if (!td) {
1089				int global = !strncmp(val, "global", 6);
1090
1091				td = get_new_job(global, &def_thread);
1092				if (!td)
1093					return 0;
1094			}
1095
1096			ret = parse_cmd_option(opt, val, options, td);
1097			if (ret) {
1098				log_err("fio: job dropped\n");
1099				put_job(td);
1100				td = NULL;
1101			}
1102			break;
1103		}
1104		default:
1105			printf("optarg <<%s>>\n", argv[optind]);
1106			break;
1107		}
1108	}
1109
1110	if (td) {
1111		ret = add_job(td, td->name ?: "fio", 0);
1112		if (ret)
1113			put_job(td);
1114	}
1115
1116	while (optind < argc) {
1117		ini_idx++;
1118		ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1119		ini_file[ini_idx - 1] = strdup(argv[optind]);
1120		optind++;
1121	}
1122
1123	return ini_idx;
1124}
1125
1126static void free_shm(void)
1127{
1128	struct shmid_ds sbuf;
1129
1130	if (threads) {
1131		shmdt((void *) threads);
1132		threads = NULL;
1133		shmctl(shm_id, IPC_RMID, &sbuf);
1134	}
1135}
1136
1137/*
1138 * The thread area is shared between the main process and the job
1139 * threads/processes. So setup a shared memory segment that will hold
1140 * all the job info.
1141 */
1142static int setup_thread_area(void)
1143{
1144	/*
1145	 * 1024 is too much on some machines, scale max_jobs if
1146	 * we get a failure that looks like too large a shm segment
1147	 */
1148	do {
1149		size_t size = max_jobs * sizeof(struct thread_data);
1150
1151		shm_id = shmget(0, size, IPC_CREAT | 0600);
1152		if (shm_id != -1)
1153			break;
1154		if (errno != EINVAL) {
1155			perror("shmget");
1156			break;
1157		}
1158
1159		max_jobs >>= 1;
1160	} while (max_jobs);
1161
1162	if (shm_id == -1)
1163		return 1;
1164
1165	threads = shmat(shm_id, NULL, 0);
1166	if (threads == (void *) -1) {
1167		perror("shmat");
1168		return 1;
1169	}
1170
1171	atexit(free_shm);
1172	return 0;
1173}
1174
1175/*
1176 * Copy the fio options into the long options map, so we mirror
1177 * job and cmd line options.
1178 */
1179static void dupe_job_options(void)
1180{
1181	struct fio_option *o;
1182	unsigned int i;
1183
1184	i = 0;
1185	while (long_options[i].name)
1186		i++;
1187
1188	o = &options[0];
1189	while (o->name) {
1190		long_options[i].name = o->name;
1191		long_options[i].val = FIO_GETOPT_JOB;
1192		if (o->type == FIO_OPT_STR_SET)
1193			long_options[i].has_arg = no_argument;
1194		else
1195			long_options[i].has_arg = required_argument;
1196
1197		i++;
1198		o++;
1199		assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1200	}
1201}
1202
1203int parse_options(int argc, char *argv[])
1204{
1205	int job_files, i;
1206
1207	f_out = stdout;
1208	f_err = stderr;
1209
1210	dupe_job_options();
1211
1212	if (setup_thread_area())
1213		return 1;
1214	if (fill_def_thread())
1215		return 1;
1216
1217	job_files = parse_cmd_line(argc, argv);
1218
1219	for (i = 0; i < job_files; i++) {
1220		if (fill_def_thread())
1221			return 1;
1222		if (parse_jobs_ini(ini_file[i], i))
1223			return 1;
1224		free(ini_file[i]);
1225	}
1226
1227	free(ini_file);
1228
1229	if (!thread_number) {
1230		log_err("No jobs defined(s)\n");
1231		return 1;
1232	}
1233
1234	return 0;
1235}
1236