init.c revision f6dea4d33fb57a6e7cb64414793c57f10b702851
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/shm.h>
13#include <sys/types.h>
14#include <sys/stat.h>
15
16#include "fio.h"
17#include "parse.h"
18#include "smalloc.h"
19#include "filehash.h"
20#include "verify.h"
21#include "profile.h"
22#include "server.h"
23
24#include "lib/getopt.h"
25
26#include "fio_version.h"
27
28#if FIO_PATCH > 0
29static char fio_version_string[] =	__fio_stringify(FIO_MAJOR) "."	\
30					__fio_stringify(FIO_MINOR) "."	\
31					__fio_stringify(FIO_PATCH);
32#else
33static char fio_version_string[] =	__fio_stringify(FIO_MAJOR) "."	\
34					__fio_stringify(FIO_MINOR);
35#endif
36
37#define FIO_RANDSEED		(0xb1899bedUL)
38
39static char **ini_file;
40static int max_jobs = FIO_MAX_JOBS;
41static int dump_cmdline;
42static int def_timeout;
43
44static struct thread_data def_thread;
45struct thread_data *threads = NULL;
46
47int exitall_on_terminate = 0;
48int terse_output = 0;
49int eta_print;
50unsigned long long mlock_size = 0;
51FILE *f_out = NULL;
52FILE *f_err = NULL;
53char **job_sections = NULL;
54int nr_job_sections = 0;
55char *exec_profile = NULL;
56int warnings_fatal = 0;
57int terse_version = 2;
58int is_backend = 0;
59int nr_clients = 0;
60int log_syslog = 0;
61
62int write_bw_log = 0;
63int read_only = 0;
64
65static int write_lat_log;
66
67static int prev_group_jobs;
68
69unsigned long fio_debug = 0;
70unsigned int fio_debug_jobno = -1;
71unsigned int *fio_debug_jobp = NULL;
72
73static char cmd_optstr[256];
74static int did_arg;
75
76const fio_fp64_t def_percentile_list[FIO_IO_U_LIST_MAX_LEN] = {
77	{ .u.f	=  1.0 },
78	{ .u.f	=  5.0 },
79	{ .u.f	= 10.0 },
80	{ .u.f	= 20.0 },
81	{ .u.f	= 30.0 },
82	{ .u.f	= 40.0 },
83	{ .u.f	= 50.0 },
84	{ .u.f	= 60.0 },
85	{ .u.f	= 70.0 },
86	{ .u.f	= 80.0 },
87	{ .u.f	= 90.0 },
88	{ .u.f	= 95.0 },
89	{ .u.f	= 99.0 },
90	{ .u.f	= 99.5 },
91	{ .u.f	= 99.9 },
92};
93
94#define FIO_CLIENT_FLAG		(1 << 16)
95
96/*
97 * Command line options. These will contain the above, plus a few
98 * extra that only pertain to fio itself and not jobs.
99 */
100static struct option l_opts[FIO_NR_OPTIONS] = {
101	{
102		.name		= (char *) "output",
103		.has_arg	= required_argument,
104		.val		= 'o' | FIO_CLIENT_FLAG,
105	},
106	{
107		.name		= (char *) "timeout",
108		.has_arg	= required_argument,
109		.val		= 't' | FIO_CLIENT_FLAG,
110	},
111	{
112		.name		= (char *) "latency-log",
113		.has_arg	= required_argument,
114		.val		= 'l' | FIO_CLIENT_FLAG,
115	},
116	{
117		.name		= (char *) "bandwidth-log",
118		.has_arg	= required_argument,
119		.val		= 'b' | FIO_CLIENT_FLAG,
120	},
121	{
122		.name		= (char *) "minimal",
123		.has_arg	= optional_argument,
124		.val		= 'm' | FIO_CLIENT_FLAG,
125	},
126	{
127		.name		= (char *) "version",
128		.has_arg	= no_argument,
129		.val		= 'v' | FIO_CLIENT_FLAG,
130	},
131	{
132		.name		= (char *) "help",
133		.has_arg	= no_argument,
134		.val		= 'h' | FIO_CLIENT_FLAG,
135	},
136	{
137		.name		= (char *) "cmdhelp",
138		.has_arg	= optional_argument,
139		.val		= 'c' | FIO_CLIENT_FLAG,
140	},
141	{
142		.name		= (char *) "showcmd",
143		.has_arg	= no_argument,
144		.val		= 's' | FIO_CLIENT_FLAG,
145	},
146	{
147		.name		= (char *) "readonly",
148		.has_arg	= no_argument,
149		.val		= 'r' | FIO_CLIENT_FLAG,
150	},
151	{
152		.name		= (char *) "eta",
153		.has_arg	= required_argument,
154		.val		= 'e' | FIO_CLIENT_FLAG,
155	},
156	{
157		.name		= (char *) "debug",
158		.has_arg	= required_argument,
159		.val		= 'd' | FIO_CLIENT_FLAG,
160	},
161	{
162		.name		= (char *) "section",
163		.has_arg	= required_argument,
164		.val		= 'x' | FIO_CLIENT_FLAG,
165	},
166	{
167		.name		= (char *) "alloc-size",
168		.has_arg	= required_argument,
169		.val		= 'a' | FIO_CLIENT_FLAG,
170	},
171	{
172		.name		= (char *) "profile",
173		.has_arg	= required_argument,
174		.val		= 'p' | FIO_CLIENT_FLAG,
175	},
176	{
177		.name		= (char *) "warnings-fatal",
178		.has_arg	= no_argument,
179		.val		= 'w' | FIO_CLIENT_FLAG,
180	},
181	{
182		.name		= (char *) "max-jobs",
183		.has_arg	= required_argument,
184		.val		= 'j' | FIO_CLIENT_FLAG,
185	},
186	{
187		.name		= (char *) "terse-version",
188		.has_arg	= required_argument,
189		.val		= 'V' | FIO_CLIENT_FLAG,
190	},
191	{
192		.name		= (char *) "server",
193		.has_arg	= optional_argument,
194		.val		= 'S',
195	},
196	{	.name		= (char *) "daemonize",
197		.has_arg	= required_argument,
198		.val		= 'D',
199	},
200	{
201		.name		= (char *) "client",
202		.has_arg	= required_argument,
203		.val		= 'C',
204	},
205	{
206		.name		= NULL,
207	},
208};
209
210static void free_shm(void)
211{
212	struct shmid_ds sbuf;
213
214	if (threads) {
215		void *tp = threads;
216
217		threads = NULL;
218		file_hash_exit();
219		fio_debug_jobp = NULL;
220		shmdt(tp);
221		shmctl(shm_id, IPC_RMID, &sbuf);
222	}
223
224	scleanup();
225}
226
227/*
228 * The thread area is shared between the main process and the job
229 * threads/processes. So setup a shared memory segment that will hold
230 * all the job info. We use the end of the region for keeping track of
231 * open files across jobs, for file sharing.
232 */
233static int setup_thread_area(void)
234{
235	void *hash;
236
237	if (threads)
238		return 0;
239
240	/*
241	 * 1024 is too much on some machines, scale max_jobs if
242	 * we get a failure that looks like too large a shm segment
243	 */
244	do {
245		size_t size = max_jobs * sizeof(struct thread_data);
246
247		size += file_hash_size;
248		size += sizeof(unsigned int);
249
250		shm_id = shmget(0, size, IPC_CREAT | 0600);
251		if (shm_id != -1)
252			break;
253		if (errno != EINVAL) {
254			perror("shmget");
255			break;
256		}
257
258		max_jobs >>= 1;
259	} while (max_jobs);
260
261	if (shm_id == -1)
262		return 1;
263
264	threads = shmat(shm_id, NULL, 0);
265	if (threads == (void *) -1) {
266		perror("shmat");
267		return 1;
268	}
269
270	memset(threads, 0, max_jobs * sizeof(struct thread_data));
271	hash = (void *) threads + max_jobs * sizeof(struct thread_data);
272	fio_debug_jobp = (void *) hash + file_hash_size;
273	*fio_debug_jobp = -1;
274	file_hash_init(hash);
275	return 0;
276}
277
278/*
279 * Return a free job structure.
280 */
281static struct thread_data *get_new_job(int global, struct thread_data *parent)
282{
283	struct thread_data *td;
284
285	if (global)
286		return &def_thread;
287	if (setup_thread_area()) {
288		log_err("error: failed to setup shm segment\n");
289		return NULL;
290	}
291	if (thread_number >= max_jobs) {
292		log_err("error: maximum number of jobs (%d) reached.\n",
293				max_jobs);
294		return NULL;
295	}
296
297	td = &threads[thread_number++];
298	*td = *parent;
299
300	td->o.uid = td->o.gid = -1U;
301
302	dup_files(td, parent);
303	options_mem_dupe(td);
304
305	profile_add_hooks(td);
306
307	td->thread_number = thread_number;
308	return td;
309}
310
311static void put_job(struct thread_data *td)
312{
313	if (td == &def_thread)
314		return;
315
316	profile_td_exit(td);
317
318	if (td->error)
319		log_info("fio: %s\n", td->verror);
320
321	memset(&threads[td->thread_number - 1], 0, sizeof(*td));
322	thread_number--;
323}
324
325static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
326{
327	unsigned int bs = td->o.min_bs[ddir];
328	unsigned long long bytes_per_sec;
329
330	assert(ddir_rw(ddir));
331
332	if (td->o.rate[ddir])
333		bytes_per_sec = td->o.rate[ddir];
334	else
335		bytes_per_sec = td->o.rate_iops[ddir] * bs;
336
337	if (!bytes_per_sec) {
338		log_err("rate lower than supported\n");
339		return -1;
340	}
341
342	td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
343	td->rate_pending_usleep[ddir] = 0;
344	return 0;
345}
346
347static int setup_rate(struct thread_data *td)
348{
349	int ret = 0;
350
351	if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
352		ret = __setup_rate(td, DDIR_READ);
353	if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
354		ret |= __setup_rate(td, DDIR_WRITE);
355
356	return ret;
357}
358
359static int fixed_block_size(struct thread_options *o)
360{
361	return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
362		o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
363		o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
364}
365
366/*
367 * Lazy way of fixing up options that depend on each other. We could also
368 * define option callback handlers, but this is easier.
369 */
370static int fixup_options(struct thread_data *td)
371{
372	struct thread_options *o = &td->o;
373	int ret = 0;
374
375#ifndef FIO_HAVE_PSHARED_MUTEX
376	if (!o->use_thread) {
377		log_info("fio: this platform does not support process shared"
378			 " mutexes, forcing use of threads. Use the 'thread'"
379			 " option to get rid of this warning.\n");
380		o->use_thread = 1;
381		ret = warnings_fatal;
382	}
383#endif
384
385	if (o->write_iolog_file && o->read_iolog_file) {
386		log_err("fio: read iolog overrides write_iolog\n");
387		free(o->write_iolog_file);
388		o->write_iolog_file = NULL;
389		ret = warnings_fatal;
390	}
391
392	/*
393	 * only really works for sequential io for now, and with 1 file
394	 */
395	if (o->zone_size && td_random(td) && o->open_files == 1)
396		o->zone_size = 0;
397
398	/*
399	 * Reads can do overwrites, we always need to pre-create the file
400	 */
401	if (td_read(td) || td_rw(td))
402		o->overwrite = 1;
403
404	if (!o->min_bs[DDIR_READ])
405		o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
406	if (!o->max_bs[DDIR_READ])
407		o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
408	if (!o->min_bs[DDIR_WRITE])
409		o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
410	if (!o->max_bs[DDIR_WRITE])
411		o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
412
413	o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
414
415	/*
416	 * For random IO, allow blockalign offset other than min_bs.
417	 */
418	if (!o->ba[DDIR_READ] || !td_random(td))
419		o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
420	if (!o->ba[DDIR_WRITE] || !td_random(td))
421		o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
422
423	if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
424	    o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
425	    !o->norandommap) {
426		log_err("fio: Any use of blockalign= turns off randommap\n");
427		o->norandommap = 1;
428		ret = warnings_fatal;
429	}
430
431	if (!o->file_size_high)
432		o->file_size_high = o->file_size_low;
433
434	if (o->norandommap && o->verify != VERIFY_NONE
435	    && !fixed_block_size(o))  {
436		log_err("fio: norandommap given for variable block sizes, "
437			"verify disabled\n");
438		o->verify = VERIFY_NONE;
439		ret = warnings_fatal;
440	}
441	if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
442		log_err("fio: bs_unaligned may not work with raw io\n");
443
444	/*
445	 * thinktime_spin must be less than thinktime
446	 */
447	if (o->thinktime_spin > o->thinktime)
448		o->thinktime_spin = o->thinktime;
449
450	/*
451	 * The low water mark cannot be bigger than the iodepth
452	 */
453	if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
454		/*
455		 * syslet work around - if the workload is sequential,
456		 * we want to let the queue drain all the way down to
457		 * avoid seeking between async threads
458		 */
459		if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
460			o->iodepth_low = 1;
461		else
462			o->iodepth_low = o->iodepth;
463	}
464
465	/*
466	 * If batch number isn't set, default to the same as iodepth
467	 */
468	if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
469		o->iodepth_batch = o->iodepth;
470
471	if (o->nr_files > td->files_index)
472		o->nr_files = td->files_index;
473
474	if (o->open_files > o->nr_files || !o->open_files)
475		o->open_files = o->nr_files;
476
477	if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
478	    ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
479		o->rate_iops_min[1]))) {
480		log_err("fio: rate and rate_iops are mutually exclusive\n");
481		ret = 1;
482	}
483	if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
484	    (o->rate_iops[0] < o->rate_iops_min[0]) ||
485	    (o->rate_iops[1] < o->rate_iops_min[1])) {
486		log_err("fio: minimum rate exceeds rate\n");
487		ret = 1;
488	}
489
490	if (!o->timeout && o->time_based) {
491		log_err("fio: time_based requires a runtime/timeout setting\n");
492		o->time_based = 0;
493		ret = warnings_fatal;
494	}
495
496	if (o->fill_device && !o->size)
497		o->size = -1ULL;
498
499	if (o->verify != VERIFY_NONE) {
500		if (td_write(td) && o->do_verify && o->numjobs > 1) {
501			log_info("Multiple writers may overwrite blocks that "
502				"belong to other jobs. This can cause "
503				"verification failures.\n");
504			ret = warnings_fatal;
505		}
506
507		o->refill_buffers = 1;
508		if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
509		    !o->verify_interval)
510			o->verify_interval = o->min_bs[DDIR_WRITE];
511	}
512
513	if (o->pre_read) {
514		o->invalidate_cache = 0;
515		if (td->io_ops->flags & FIO_PIPEIO) {
516			log_info("fio: cannot pre-read files with an IO engine"
517				 " that isn't seekable. Pre-read disabled.\n");
518			ret = warnings_fatal;
519		}
520	}
521
522#ifndef FIO_HAVE_FDATASYNC
523	if (o->fdatasync_blocks) {
524		log_info("fio: this platform does not support fdatasync()"
525			 " falling back to using fsync().  Use the 'fsync'"
526			 " option instead of 'fdatasync' to get rid of"
527			 " this warning\n");
528		o->fsync_blocks = o->fdatasync_blocks;
529		o->fdatasync_blocks = 0;
530		ret = warnings_fatal;
531	}
532#endif
533
534	return ret;
535}
536
537/*
538 * This function leaks the buffer
539 */
540static char *to_kmg(unsigned int val)
541{
542	char *buf = malloc(32);
543	char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
544	char *p = post;
545
546	do {
547		if (val & 1023)
548			break;
549
550		val >>= 10;
551		p++;
552	} while (*p);
553
554	snprintf(buf, 31, "%u%c", val, *p);
555	return buf;
556}
557
558/* External engines are specified by "external:name.o") */
559static const char *get_engine_name(const char *str)
560{
561	char *p = strstr(str, ":");
562
563	if (!p)
564		return str;
565
566	p++;
567	strip_blank_front(&p);
568	strip_blank_end(p);
569	return p;
570}
571
572static int exists_and_not_file(const char *filename)
573{
574	struct stat sb;
575
576	if (lstat(filename, &sb) == -1)
577		return 0;
578
579	/* \\.\ is the device namespace in Windows, where every file
580	 * is a device node */
581	if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
582		return 0;
583
584	return 1;
585}
586
587static void td_fill_rand_seeds_os(struct thread_data *td)
588{
589	os_random_seed(td->rand_seeds[0], &td->bsrange_state);
590	os_random_seed(td->rand_seeds[1], &td->verify_state);
591	os_random_seed(td->rand_seeds[2], &td->rwmix_state);
592
593	if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
594		os_random_seed(td->rand_seeds[3], &td->next_file_state);
595
596	os_random_seed(td->rand_seeds[5], &td->file_size_state);
597	os_random_seed(td->rand_seeds[6], &td->trim_state);
598
599	if (!td_random(td))
600		return;
601
602	if (td->o.rand_repeatable)
603		td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
604
605	os_random_seed(td->rand_seeds[4], &td->random_state);
606}
607
608static void td_fill_rand_seeds_internal(struct thread_data *td)
609{
610	init_rand_seed(&td->__bsrange_state, td->rand_seeds[0]);
611	init_rand_seed(&td->__verify_state, td->rand_seeds[1]);
612	init_rand_seed(&td->__rwmix_state, td->rand_seeds[2]);
613
614	if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
615		init_rand_seed(&td->__next_file_state, td->rand_seeds[3]);
616
617	init_rand_seed(&td->__file_size_state, td->rand_seeds[5]);
618	init_rand_seed(&td->__trim_state, td->rand_seeds[6]);
619
620	if (!td_random(td))
621		return;
622
623	if (td->o.rand_repeatable)
624		td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
625
626	init_rand_seed(&td->__random_state, td->rand_seeds[4]);
627}
628
629void td_fill_rand_seeds(struct thread_data *td)
630{
631	if (td->o.use_os_rand)
632		td_fill_rand_seeds_os(td);
633	else
634		td_fill_rand_seeds_internal(td);
635
636	init_rand_seed(&td->buf_state, td->rand_seeds[7]);
637}
638
639/*
640 * Initialize the various random states we need (random io, block size ranges,
641 * read/write mix, etc).
642 */
643static int init_random_state(struct thread_data *td)
644{
645	int fd;
646
647	fd = open("/dev/urandom", O_RDONLY);
648	if (fd == -1) {
649		td_verror(td, errno, "open");
650		return 1;
651	}
652
653	if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
654	    (int) sizeof(td->rand_seeds)) {
655		td_verror(td, EIO, "read");
656		close(fd);
657		return 1;
658	}
659
660	close(fd);
661	td_fill_rand_seeds(td);
662	return 0;
663}
664
665/*
666 * Adds a job to the list of things todo. Sanitizes the various options
667 * to make sure we don't have conflicts, and initializes various
668 * members of td.
669 */
670static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
671{
672	const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
673				   "randread", "randwrite", "randrw" };
674	unsigned int i;
675	const char *engine;
676	char fname[PATH_MAX];
677	int numjobs, file_alloced;
678
679	/*
680	 * the def_thread is just for options, it's not a real job
681	 */
682	if (td == &def_thread)
683		return 0;
684
685	/*
686	 * if we are just dumping the output command line, don't add the job
687	 */
688	if (dump_cmdline) {
689		put_job(td);
690		return 0;
691	}
692
693	if (profile_td_init(td))
694		goto err;
695
696	engine = get_engine_name(td->o.ioengine);
697	td->io_ops = load_ioengine(td, engine);
698	if (!td->io_ops) {
699		log_err("fio: failed to load engine %s\n", engine);
700		goto err;
701	}
702
703	if (td->o.use_thread)
704		nr_thread++;
705	else
706		nr_process++;
707
708	if (td->o.odirect)
709		td->io_ops->flags |= FIO_RAWIO;
710
711	file_alloced = 0;
712	if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
713		file_alloced = 1;
714
715		if (td->o.nr_files == 1 && exists_and_not_file(jobname))
716			add_file(td, jobname);
717		else {
718			for (i = 0; i < td->o.nr_files; i++) {
719				sprintf(fname, "%s.%d.%d", jobname,
720							td->thread_number, i);
721				add_file(td, fname);
722			}
723		}
724	}
725
726	if (fixup_options(td))
727		goto err;
728
729	if (td->io_ops->flags & FIO_DISKLESSIO) {
730		struct fio_file *f;
731
732		for_each_file(td, f, i)
733			f->real_file_size = -1ULL;
734	}
735
736	td->mutex = fio_mutex_init(0);
737
738	td->ts.clat_percentiles = td->o.clat_percentiles;
739	if (td->o.overwrite_plist)
740		memcpy(td->ts.percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
741	else
742		memcpy(td->ts.percentile_list, def_percentile_list, sizeof(def_percentile_list));
743
744	td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
745	td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
746	td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
747	td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
748	td->ddir_seq_nr = td->o.ddir_seq_nr;
749
750	if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
751		prev_group_jobs = 0;
752		groupid++;
753	}
754
755	td->groupid = groupid;
756	prev_group_jobs++;
757
758	if (init_random_state(td))
759		goto err;
760
761	if (setup_rate(td))
762		goto err;
763
764	if (td->o.write_lat_log) {
765		setup_log(&td->lat_log);
766		setup_log(&td->slat_log);
767		setup_log(&td->clat_log);
768	}
769	if (td->o.write_bw_log)
770		setup_log(&td->bw_log);
771	if (td->o.write_iops_log)
772		setup_log(&td->iops_log);
773
774	if (!td->o.name)
775		td->o.name = strdup(jobname);
776
777	if (!terse_output) {
778		if (!job_add_num) {
779			if (!strcmp(td->io_ops->name, "cpuio")) {
780				log_info("%s: ioengine=cpu, cpuload=%u,"
781					 " cpucycle=%u\n", td->o.name,
782							td->o.cpuload,
783							td->o.cpucycle);
784			} else {
785				char *c1, *c2, *c3, *c4;
786
787				c1 = to_kmg(td->o.min_bs[DDIR_READ]);
788				c2 = to_kmg(td->o.max_bs[DDIR_READ]);
789				c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
790				c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
791
792				log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
793					 " ioengine=%s, iodepth=%u\n",
794						td->o.name, td->groupid,
795						ddir_str[td->o.td_ddir],
796						c1, c2, c3, c4,
797						td->io_ops->name,
798						td->o.iodepth);
799
800				free(c1);
801				free(c2);
802				free(c3);
803				free(c4);
804			}
805		} else if (job_add_num == 1)
806			log_info("...\n");
807	}
808
809	/*
810	 * recurse add identical jobs, clear numjobs and stonewall options
811	 * as they don't apply to sub-jobs
812	 */
813	numjobs = td->o.numjobs;
814	while (--numjobs) {
815		struct thread_data *td_new = get_new_job(0, td);
816
817		if (!td_new)
818			goto err;
819
820		td_new->o.numjobs = 1;
821		td_new->o.stonewall = 0;
822		td_new->o.new_group = 0;
823
824		if (file_alloced) {
825			td_new->o.filename = NULL;
826			td_new->files_index = 0;
827			td_new->files_size = 0;
828			td_new->files = NULL;
829		}
830
831		job_add_num = numjobs - 1;
832
833		if (add_job(td_new, jobname, job_add_num))
834			goto err;
835	}
836
837	return 0;
838err:
839	put_job(td);
840	return -1;
841}
842
843/*
844 * Parse as if 'o' was a command line
845 */
846void add_job_opts(const char **o)
847{
848	struct thread_data *td, *td_parent;
849	int i, in_global = 1;
850	char jobname[32];
851
852	i = 0;
853	td_parent = td = NULL;
854	while (o[i]) {
855		if (!strncmp(o[i], "name", 4)) {
856			in_global = 0;
857			if (td)
858				add_job(td, jobname, 0);
859			td = NULL;
860			sprintf(jobname, "%s", o[i] + 5);
861		}
862		if (in_global && !td_parent)
863			td_parent = get_new_job(1, &def_thread);
864		else if (!in_global && !td) {
865			if (!td_parent)
866				td_parent = &def_thread;
867			td = get_new_job(0, td_parent);
868		}
869		if (in_global)
870			fio_options_parse(td_parent, (char **) &o[i], 1);
871		else
872			fio_options_parse(td, (char **) &o[i], 1);
873		i++;
874	}
875
876	if (td)
877		add_job(td, jobname, 0);
878}
879
880static int skip_this_section(const char *name)
881{
882	int i;
883
884	if (!nr_job_sections)
885		return 0;
886	if (!strncmp(name, "global", 6))
887		return 0;
888
889	for (i = 0; i < nr_job_sections; i++)
890		if (!strcmp(job_sections[i], name))
891			return 0;
892
893	return 1;
894}
895
896static int is_empty_or_comment(char *line)
897{
898	unsigned int i;
899
900	for (i = 0; i < strlen(line); i++) {
901		if (line[i] == ';')
902			return 1;
903		if (line[i] == '#')
904			return 1;
905		if (!isspace((int) line[i]) && !iscntrl((int) line[i]))
906			return 0;
907	}
908
909	return 1;
910}
911
912/*
913 * This is our [ini] type file parser.
914 */
915int parse_jobs_ini(char *file, int is_buf, int stonewall_flag)
916{
917	unsigned int global;
918	struct thread_data *td;
919	char *string, *name;
920	FILE *f;
921	char *p;
922	int ret = 0, stonewall;
923	int first_sect = 1;
924	int skip_fgets = 0;
925	int inside_skip = 0;
926	char **opts;
927	int i, alloc_opts, num_opts;
928
929	if (is_buf)
930		f = NULL;
931	else {
932		if (!strcmp(file, "-"))
933			f = stdin;
934		else
935			f = fopen(file, "r");
936
937		if (!f) {
938			perror("fopen job file");
939			return 1;
940		}
941	}
942
943	string = malloc(4096);
944
945	/*
946	 * it's really 256 + small bit, 280 should suffice
947	 */
948	name = malloc(280);
949	memset(name, 0, 280);
950
951	alloc_opts = 8;
952	opts = malloc(sizeof(char *) * alloc_opts);
953	num_opts = 0;
954
955	stonewall = stonewall_flag;
956	do {
957		/*
958		 * if skip_fgets is set, we already have loaded a line we
959		 * haven't handled.
960		 */
961		if (!skip_fgets) {
962			if (is_buf)
963				p = strsep(&file, "\n");
964			else
965				p = fgets(string, 4095, f);
966			if (!p)
967				break;
968		}
969
970		skip_fgets = 0;
971		strip_blank_front(&p);
972		strip_blank_end(p);
973
974		if (is_empty_or_comment(p))
975			continue;
976		if (sscanf(p, "[%255[^\n]]", name) != 1) {
977			if (inside_skip)
978				continue;
979			log_err("fio: option <%s> outside of [] job section\n",
980									p);
981			break;
982		}
983
984		name[strlen(name) - 1] = '\0';
985
986		if (skip_this_section(name)) {
987			inside_skip = 1;
988			continue;
989		} else
990			inside_skip = 0;
991
992		global = !strncmp(name, "global", 6);
993
994		if (dump_cmdline) {
995			if (first_sect)
996				log_info("fio ");
997			if (!global)
998				log_info("--name=%s ", name);
999			first_sect = 0;
1000		}
1001
1002		td = get_new_job(global, &def_thread);
1003		if (!td) {
1004			ret = 1;
1005			break;
1006		}
1007
1008		/*
1009		 * Seperate multiple job files by a stonewall
1010		 */
1011		if (!global && stonewall) {
1012			td->o.stonewall = stonewall;
1013			stonewall = 0;
1014		}
1015
1016		num_opts = 0;
1017		memset(opts, 0, alloc_opts * sizeof(char *));
1018
1019		while (1) {
1020			if (is_buf)
1021				p = strsep(&file, "\n");
1022			else
1023				p = fgets(string, 4096, f);
1024			if (!p)
1025				break;
1026
1027			if (is_empty_or_comment(p))
1028				continue;
1029
1030			strip_blank_front(&p);
1031
1032			/*
1033			 * new section, break out and make sure we don't
1034			 * fgets() a new line at the top.
1035			 */
1036			if (p[0] == '[') {
1037				skip_fgets = 1;
1038				break;
1039			}
1040
1041			strip_blank_end(p);
1042
1043			if (num_opts == alloc_opts) {
1044				alloc_opts <<= 1;
1045				opts = realloc(opts,
1046						alloc_opts * sizeof(char *));
1047			}
1048
1049			opts[num_opts] = strdup(p);
1050			num_opts++;
1051		}
1052
1053		ret = fio_options_parse(td, opts, num_opts);
1054		if (!ret) {
1055			if (dump_cmdline)
1056				for (i = 0; i < num_opts; i++)
1057					log_info("--%s ", opts[i]);
1058
1059			ret = add_job(td, name, 0);
1060		} else {
1061			log_err("fio: job %s dropped\n", name);
1062			put_job(td);
1063		}
1064
1065		for (i = 0; i < num_opts; i++)
1066			free(opts[i]);
1067		num_opts = 0;
1068	} while (!ret);
1069
1070	if (dump_cmdline)
1071		log_info("\n");
1072
1073	for (i = 0; i < num_opts; i++)
1074		free(opts[i]);
1075
1076	free(string);
1077	free(name);
1078	free(opts);
1079	if (!is_buf && f != stdin)
1080		fclose(f);
1081	return ret;
1082}
1083
1084static int fill_def_thread(void)
1085{
1086	memset(&def_thread, 0, sizeof(def_thread));
1087
1088	fio_getaffinity(getpid(), &def_thread.o.cpumask);
1089	def_thread.o.timeout = def_timeout;
1090
1091	/*
1092	 * fill default options
1093	 */
1094	fio_fill_default_options(&def_thread);
1095	return 0;
1096}
1097
1098static void usage(const char *name)
1099{
1100	printf("fio %s\n", fio_version_string);
1101	printf("%s [options] [job options] <job file(s)>\n", name);
1102	printf("\t--debug=options\tEnable debug logging\n");
1103	printf("\t--output\tWrite output to file\n");
1104	printf("\t--timeout\tRuntime in seconds\n");
1105	printf("\t--latency-log\tGenerate per-job latency logs\n");
1106	printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1107	printf("\t--minimal\tMinimal (terse) output\n");
1108	printf("\t--version\tPrint version info and exit\n");
1109	printf("\t--terse-version=x Terse version output format\n");
1110	printf("\t--help\t\tPrint this page\n");
1111	printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
1112		" them\n");
1113	printf("\t--showcmd\tTurn a job file into command line options\n");
1114	printf("\t--eta=when\tWhen ETA estimate should be printed\n");
1115	printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
1116	printf("\t--readonly\tTurn on safety read-only checks, preventing"
1117		" writes\n");
1118	printf("\t--section=name\tOnly run specified section in job file\n");
1119	printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
1120		" (def 1024)\n");
1121	printf("\t--warnings-fatal Fio parser warnings are fatal\n");
1122	printf("\t--max-jobs\tMaximum number of threads/processes to support\n");
1123	printf("\t--server=args\tStart a backend fio server\n");
1124	printf("\t--daemonize=pidfile Background fio server, write pid to file\n");
1125	printf("\t--client=hostname Talk to remote backend fio server at hostname\n");
1126	printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
1127	printf("\n                   Jens Axboe <jaxboe@fusionio.com>\n");
1128}
1129
1130#ifdef FIO_INC_DEBUG
1131struct debug_level debug_levels[] = {
1132	{ .name = "process",	.shift = FD_PROCESS, },
1133	{ .name = "file",	.shift = FD_FILE, },
1134	{ .name = "io",		.shift = FD_IO, },
1135	{ .name = "mem",	.shift = FD_MEM, },
1136	{ .name = "blktrace",	.shift = FD_BLKTRACE },
1137	{ .name = "verify",	.shift = FD_VERIFY },
1138	{ .name = "random",	.shift = FD_RANDOM },
1139	{ .name = "parse",	.shift = FD_PARSE },
1140	{ .name = "diskutil",	.shift = FD_DISKUTIL },
1141	{ .name = "job",	.shift = FD_JOB },
1142	{ .name = "mutex",	.shift = FD_MUTEX },
1143	{ .name	= "profile",	.shift = FD_PROFILE },
1144	{ .name = "time",	.shift = FD_TIME },
1145	{ .name = "net",	.shift = FD_NET },
1146	{ .name = NULL, },
1147};
1148
1149static int set_debug(const char *string)
1150{
1151	struct debug_level *dl;
1152	char *p = (char *) string;
1153	char *opt;
1154	int i;
1155
1156	if (!strcmp(string, "?") || !strcmp(string, "help")) {
1157		log_info("fio: dumping debug options:");
1158		for (i = 0; debug_levels[i].name; i++) {
1159			dl = &debug_levels[i];
1160			log_info("%s,", dl->name);
1161		}
1162		log_info("all\n");
1163		return 1;
1164	}
1165
1166	while ((opt = strsep(&p, ",")) != NULL) {
1167		int found = 0;
1168
1169		if (!strncmp(opt, "all", 3)) {
1170			log_info("fio: set all debug options\n");
1171			fio_debug = ~0UL;
1172			continue;
1173		}
1174
1175		for (i = 0; debug_levels[i].name; i++) {
1176			dl = &debug_levels[i];
1177			found = !strncmp(opt, dl->name, strlen(dl->name));
1178			if (!found)
1179				continue;
1180
1181			if (dl->shift == FD_JOB) {
1182				opt = strchr(opt, ':');
1183				if (!opt) {
1184					log_err("fio: missing job number\n");
1185					break;
1186				}
1187				opt++;
1188				fio_debug_jobno = atoi(opt);
1189				log_info("fio: set debug jobno %d\n",
1190							fio_debug_jobno);
1191			} else {
1192				log_info("fio: set debug option %s\n", opt);
1193				fio_debug |= (1UL << dl->shift);
1194			}
1195			break;
1196		}
1197
1198		if (!found)
1199			log_err("fio: debug mask %s not found\n", opt);
1200	}
1201	return 0;
1202}
1203#else
1204static int set_debug(const char *string)
1205{
1206	log_err("fio: debug tracing not included in build\n");
1207	return 1;
1208}
1209#endif
1210
1211static void fio_options_fill_optstring(void)
1212{
1213	char *ostr = cmd_optstr;
1214	int i, c;
1215
1216	c = i = 0;
1217	while (l_opts[i].name) {
1218		ostr[c++] = l_opts[i].val;
1219		if (l_opts[i].has_arg == required_argument)
1220			ostr[c++] = ':';
1221		else if (l_opts[i].has_arg == optional_argument) {
1222			ostr[c++] = ':';
1223			ostr[c++] = ':';
1224		}
1225		i++;
1226	}
1227	ostr[c] = '\0';
1228}
1229
1230static int client_flag_set(char c)
1231{
1232	int i;
1233
1234	i = 0;
1235	while (l_opts[i].name) {
1236		int val = l_opts[i].val;
1237
1238		if (c == (val & 0xff))
1239			return (val & FIO_CLIENT_FLAG);
1240
1241		i++;
1242	}
1243
1244	return 0;
1245}
1246
1247void parse_cmd_client(void *client, char *opt)
1248{
1249	fio_client_add_cmd_option(client, opt);
1250}
1251
1252int parse_cmd_line(int argc, char *argv[])
1253{
1254	struct thread_data *td = NULL;
1255	int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1256	char *ostr = cmd_optstr;
1257	void *pid_file = NULL;
1258	void *cur_client = NULL;
1259	int backend = 0;
1260
1261	/*
1262	 * Reset optind handling, since we may call this multiple times
1263	 * for the backend.
1264	 */
1265	optind = 1;
1266
1267	while ((c = getopt_long_only(argc, argv, ostr, l_opts, &lidx)) != -1) {
1268		did_arg = 1;
1269
1270		if ((c & FIO_CLIENT_FLAG) || client_flag_set(c)) {
1271			parse_cmd_client(cur_client, argv[optind - 1]);
1272			c &= ~FIO_CLIENT_FLAG;
1273		}
1274
1275		switch (c) {
1276		case 'a':
1277			smalloc_pool_size = atoi(optarg);
1278			break;
1279		case 't':
1280			def_timeout = atoi(optarg);
1281			break;
1282		case 'l':
1283			write_lat_log = 1;
1284			break;
1285		case 'b':
1286			write_bw_log = 1;
1287			break;
1288		case 'o':
1289			f_out = fopen(optarg, "w+");
1290			if (!f_out) {
1291				perror("fopen output");
1292				exit(1);
1293			}
1294			f_err = f_out;
1295			break;
1296		case 'm':
1297			terse_output = 1;
1298			break;
1299		case 'h':
1300			if (!cur_client) {
1301				usage(argv[0]);
1302				do_exit++;
1303			}
1304			break;
1305		case 'c':
1306			if (!cur_client) {
1307				fio_show_option_help(optarg);
1308				do_exit++;
1309			}
1310			break;
1311		case 's':
1312			dump_cmdline = 1;
1313			break;
1314		case 'r':
1315			read_only = 1;
1316			break;
1317		case 'v':
1318			if (!cur_client) {
1319				log_info("fio %s\n", fio_version_string);
1320				do_exit++;
1321			}
1322			break;
1323		case 'V':
1324			terse_version = atoi(optarg);
1325			if (terse_version != 3) {
1326				log_err("fio: bad terse version format\n");
1327				exit_val = 1;
1328				do_exit++;
1329			}
1330			break;
1331		case 'e':
1332			if (!strcmp("always", optarg))
1333				eta_print = FIO_ETA_ALWAYS;
1334			else if (!strcmp("never", optarg))
1335				eta_print = FIO_ETA_NEVER;
1336			break;
1337		case 'd':
1338			if (set_debug(optarg))
1339				do_exit++;
1340			break;
1341		case 'x': {
1342			size_t new_size;
1343
1344			if (!strcmp(optarg, "global")) {
1345				log_err("fio: can't use global as only "
1346					"section\n");
1347				do_exit++;
1348				exit_val = 1;
1349				break;
1350			}
1351			new_size = (nr_job_sections + 1) * sizeof(char *);
1352			job_sections = realloc(job_sections, new_size);
1353			job_sections[nr_job_sections] = strdup(optarg);
1354			nr_job_sections++;
1355			break;
1356			}
1357		case 'p':
1358			exec_profile = strdup(optarg);
1359			break;
1360		case FIO_GETOPT_JOB: {
1361			const char *opt = l_opts[lidx].name;
1362			char *val = optarg;
1363
1364			if (!strncmp(opt, "name", 4) && td) {
1365				ret = add_job(td, td->o.name ?: "fio", 0);
1366				if (ret)
1367					return 0;
1368				td = NULL;
1369			}
1370			if (!td) {
1371				int is_section = !strncmp(opt, "name", 4);
1372				int global = 0;
1373
1374				if (!is_section || !strncmp(val, "global", 6))
1375					global = 1;
1376
1377				if (is_section && skip_this_section(val))
1378					continue;
1379
1380				td = get_new_job(global, &def_thread);
1381				if (!td)
1382					return 0;
1383			}
1384
1385			ret = fio_cmd_option_parse(td, opt, val);
1386			break;
1387		}
1388		case 'w':
1389			warnings_fatal = 1;
1390			break;
1391		case 'j':
1392			max_jobs = atoi(optarg);
1393			if (!max_jobs || max_jobs > REAL_MAX_JOBS) {
1394				log_err("fio: invalid max jobs: %d\n", max_jobs);
1395				do_exit++;
1396				exit_val = 1;
1397			}
1398			break;
1399		case 'S':
1400			if (nr_clients) {
1401				log_err("fio: can't be both client and server\n");
1402				do_exit++;
1403				exit_val = 1;
1404				break;
1405			}
1406			if (optarg)
1407				fio_server_set_arg(optarg);
1408			is_backend = 1;
1409			backend = 1;
1410			break;
1411		case 'D':
1412			pid_file = strdup(optarg);
1413			break;
1414		case 'C':
1415			if (is_backend) {
1416				log_err("fio: can't be both client and server\n");
1417				do_exit++;
1418				exit_val = 1;
1419				break;
1420			}
1421			if (fio_client_add(optarg, &cur_client)) {
1422				log_err("fio: failed adding client %s\n", optarg);
1423				do_exit++;
1424				exit_val = 1;
1425				break;
1426			}
1427			break;
1428		default:
1429			do_exit++;
1430			exit_val = 1;
1431			break;
1432		}
1433		if (do_exit)
1434			break;
1435	}
1436
1437	if (do_exit) {
1438		if (exit_val && !(is_backend || nr_clients))
1439			exit(exit_val);
1440	}
1441
1442	if (nr_clients && fio_clients_connect()) {
1443		do_exit++;
1444		exit_val = 1;
1445		return -1;
1446	}
1447
1448	if (is_backend && backend)
1449		return fio_start_server(pid_file);
1450
1451	if (td) {
1452		if (!ret)
1453			ret = add_job(td, td->o.name ?: "fio", 0);
1454	}
1455
1456	while (!ret && optind < argc) {
1457		ini_idx++;
1458		ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1459		ini_file[ini_idx - 1] = strdup(argv[optind]);
1460		optind++;
1461	}
1462
1463	return ini_idx;
1464}
1465
1466int parse_options(int argc, char *argv[])
1467{
1468	int job_files, i;
1469
1470	f_out = stdout;
1471	f_err = stderr;
1472
1473	fio_options_fill_optstring();
1474	fio_options_dup_and_init(l_opts);
1475
1476	atexit(free_shm);
1477
1478	if (fill_def_thread())
1479		return 1;
1480
1481	job_files = parse_cmd_line(argc, argv);
1482
1483	if (job_files > 0) {
1484		for (i = 0; i < job_files; i++) {
1485			if (fill_def_thread())
1486				return 1;
1487			if (nr_clients) {
1488				if (fio_clients_send_ini(ini_file[i]))
1489					return 1;
1490				free(ini_file[i]);
1491			} else if (!is_backend) {
1492				if (parse_jobs_ini(ini_file[i], 0, i))
1493					return 1;
1494				free(ini_file[i]);
1495			}
1496		}
1497	}
1498
1499	free(ini_file);
1500	options_mem_free(&def_thread);
1501
1502	if (!thread_number) {
1503		if (dump_cmdline)
1504			return 0;
1505		if (exec_profile)
1506			return 0;
1507		if (is_backend || nr_clients)
1508			return 0;
1509		if (did_arg)
1510			return 0;
1511
1512		log_err("No jobs(s) defined\n\n");
1513
1514		if (!did_arg) {
1515			usage(argv[0]);
1516			return 1;
1517		}
1518
1519		return 0;
1520	}
1521
1522	if (def_thread.o.gtod_offload) {
1523		fio_gtod_init();
1524		fio_gtod_offload = 1;
1525		fio_gtod_cpu = def_thread.o.gtod_cpu;
1526	}
1527
1528	if (!terse_output)
1529		log_info("fio %s\n", fio_version_string);
1530
1531	return 0;
1532}
1533