init.c revision e6203e9e73039537d41810e91e3dfe6abeb919d9 
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
20static char fio_version_string[] = "fio 1.17.1";
21
22#define FIO_RANDSEED		(0xb1899bedUL)
23
24static char **ini_file;
25static int max_jobs = MAX_JOBS;
26static int dump_cmdline;
27
28struct thread_data def_thread;
29struct thread_data *threads = NULL;
30
31int exitall_on_terminate = 0;
32int terse_output = 0;
33unsigned long long mlock_size = 0;
34FILE *f_out = NULL;
35FILE *f_err = NULL;
36
37int write_bw_log = 0;
38
39static int def_timeout = 0;
40static int write_lat_log = 0;
41
42static int prev_group_jobs;
43
44/*
45 * Command line options. These will contain the above, plus a few
46 * extra that only pertain to fio itself and not jobs.
47 */
48static struct option long_options[FIO_NR_OPTIONS] = {
49	{
50		.name		= "output",
51		.has_arg	= required_argument,
52		.val		= 'o',
53	},
54	{
55		.name		= "timeout",
56		.has_arg	= required_argument,
57		.val		= 't',
58	},
59	{
60		.name		= "latency-log",
61		.has_arg	= required_argument,
62		.val		= 'l',
63	},
64	{
65		.name		= "bandwidth-log",
66		.has_arg	= required_argument,
67		.val		= 'b',
68	},
69	{
70		.name		= "minimal",
71		.has_arg	= optional_argument,
72		.val		= 'm',
73	},
74	{
75		.name		= "version",
76		.has_arg	= no_argument,
77		.val		= 'v',
78	},
79	{
80		.name		= "help",
81		.has_arg	= no_argument,
82		.val		= 'h',
83	},
84	{
85		.name		= "cmdhelp",
86		.has_arg	= optional_argument,
87		.val		= 'c',
88	},
89	{
90		.name		= "showcmd",
91		.has_arg	= no_argument,
92		.val		= 's'
93	},
94	{
95		.name		= NULL,
96	},
97};
98
99FILE *get_f_out()
100{
101	return f_out;
102}
103
104FILE *get_f_err()
105{
106	return f_err;
107}
108
109/*
110 * Return a free job structure.
111 */
112static struct thread_data *get_new_job(int global, struct thread_data *parent)
113{
114	struct thread_data *td;
115
116	if (global)
117		return &def_thread;
118	if (thread_number >= max_jobs)
119		return NULL;
120
121	td = &threads[thread_number++];
122	*td = *parent;
123
124	dup_files(td, parent);
125	options_mem_dupe(td);
126
127	td->thread_number = thread_number;
128	return td;
129}
130
131static void put_job(struct thread_data *td)
132{
133	if (td == &def_thread)
134		return;
135
136	if (td->error)
137		log_info("fio: %s\n", td->verror);
138
139	memset(&threads[td->thread_number - 1], 0, sizeof(*td));
140	thread_number--;
141}
142
143static int setup_rate(struct thread_data *td)
144{
145	unsigned long nr_reads_per_msec;
146	unsigned long long rate;
147	unsigned int bs;
148
149	if (!td->o.rate && !td->o.rate_iops)
150		return 0;
151
152	if (td_rw(td))
153		bs = td->o.rw_min_bs;
154	else if (td_read(td))
155		bs = td->o.min_bs[DDIR_READ];
156	else
157		bs = td->o.min_bs[DDIR_WRITE];
158
159	if (td->o.rate) {
160		rate = td->o.rate;
161		nr_reads_per_msec = (rate * 1024 * 1000LL) / bs;
162	} else
163		nr_reads_per_msec = td->o.rate_iops * 1000UL;
164
165	if (!nr_reads_per_msec) {
166		log_err("rate lower than supported\n");
167		return -1;
168	}
169
170	td->rate_usec_cycle = 1000000000ULL / nr_reads_per_msec;
171	td->rate_pending_usleep = 0;
172	return 0;
173}
174
175/*
176 * Lazy way of fixing up options that depend on each other. We could also
177 * define option callback handlers, but this is easier.
178 */
179static int fixup_options(struct thread_data *td)
180{
181	struct thread_options *o = &td->o;
182
183	if (o->rwmix[DDIR_READ] + o->rwmix[DDIR_WRITE] > 100)
184		o->rwmix[DDIR_WRITE] = 100 - o->rwmix[DDIR_READ];
185
186	if (o->write_iolog_file && o->read_iolog_file) {
187		log_err("fio: read iolog overrides write_iolog\n");
188		free(o->write_iolog_file);
189		o->write_iolog_file = NULL;
190	}
191
192	if (td->io_ops->flags & FIO_SYNCIO)
193		o->iodepth = 1;
194	else {
195		if (!o->iodepth)
196			o->iodepth = o->open_files;
197	}
198
199	/*
200	 * only really works for sequential io for now, and with 1 file
201	 */
202	if (o->zone_size && td_random(td) && o->open_files == 1)
203		o->zone_size = 0;
204
205	/*
206	 * Reads can do overwrites, we always need to pre-create the file
207	 */
208	if (td_read(td) || td_rw(td))
209		o->overwrite = 1;
210
211	if (!o->min_bs[DDIR_READ])
212		o->min_bs[DDIR_READ]= o->bs[DDIR_READ];
213	if (!o->max_bs[DDIR_READ])
214		o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
215	if (!o->min_bs[DDIR_WRITE])
216		o->min_bs[DDIR_WRITE]= o->bs[DDIR_WRITE];
217	if (!o->max_bs[DDIR_WRITE])
218		o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
219
220	o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
221
222	if (!o->file_size_high)
223		o->file_size_high = o->file_size_low;
224
225	if (o->norandommap && o->verify != VERIFY_NONE) {
226		log_err("fio: norandommap given, verify disabled\n");
227		o->verify = VERIFY_NONE;
228	}
229	if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
230		log_err("fio: bs_unaligned may not work with raw io\n");
231
232	/*
233	 * thinktime_spin must be less than thinktime
234	 */
235	if (o->thinktime_spin > o->thinktime)
236		o->thinktime_spin = o->thinktime;
237
238	/*
239	 * The low water mark cannot be bigger than the iodepth
240	 */
241	if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
242		/*
243		 * syslet work around - if the workload is sequential,
244		 * we want to let the queue drain all the way down to
245		 * avoid seeking between async threads
246		 */
247		if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
248			o->iodepth_low = 1;
249		else
250			o->iodepth_low = o->iodepth;
251	}
252
253	/*
254	 * If batch number isn't set, default to the same as iodepth
255	 */
256	if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
257		o->iodepth_batch = o->iodepth;
258
259	if (o->nr_files > td->files_index)
260		o->nr_files = td->files_index;
261
262	if (o->open_files > o->nr_files || !o->open_files)
263		o->open_files = o->nr_files;
264
265	if ((o->rate && o->rate_iops) || (o->ratemin && o->rate_iops_min)) {
266		log_err("fio: rate and rate_iops are mutually exclusive\n");
267		return 1;
268	}
269	if ((o->rate < o->ratemin) || (o->rate_iops < o->rate_iops_min)) {
270		log_err("fio: minimum rate exceeds rate\n");
271		return 1;
272	}
273
274	if (!o->timeout && o->time_based) {
275		log_err("fio: time_based requires a runtime/timeout setting\n");
276		o->time_based = 0;
277	}
278
279	return 0;
280}
281
282/*
283 * This function leaks the buffer
284 */
285static char *to_kmg(unsigned int val)
286{
287	char *buf = malloc(32);
288	char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
289	char *p = post;
290
291	do {
292		if (val & 1023)
293			break;
294
295		val >>= 10;
296		p++;
297	} while (*p);
298
299	snprintf(buf, 31, "%u%c", val, *p);
300	return buf;
301}
302
303/* External engines are specified by "external:name.o") */
304static const char *get_engine_name(const char *str)
305{
306	char *p = strstr(str, ":");
307
308	if (!p)
309		return str;
310
311	p++;
312	strip_blank_front(&p);
313	strip_blank_end(p);
314	return p;
315}
316
317static int exists_and_not_file(const char *filename)
318{
319	struct stat sb;
320
321	if (lstat(filename, &sb) == -1)
322		return 0;
323
324	if (S_ISREG(sb.st_mode))
325		return 0;
326
327	return 1;
328}
329
330/*
331 * Initialize the various random states we need (random io, block size ranges,
332 * read/write mix, etc).
333 */
334static int init_random_state(struct thread_data *td)
335{
336	unsigned long seeds[6];
337	int fd;
338
339	fd = open("/dev/urandom", O_RDONLY);
340	if (fd == -1) {
341		td_verror(td, errno, "open");
342		return 1;
343	}
344
345	if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
346		td_verror(td, EIO, "read");
347		close(fd);
348		return 1;
349	}
350
351	close(fd);
352
353	os_random_seed(seeds[0], &td->bsrange_state);
354	os_random_seed(seeds[1], &td->verify_state);
355	os_random_seed(seeds[2], &td->rwmix_state);
356
357	if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
358		os_random_seed(seeds[3], &td->next_file_state);
359
360	os_random_seed(seeds[5], &td->file_size_state);
361
362	if (!td_random(td))
363		return 0;
364
365	if (td->o.rand_repeatable)
366		seeds[4] = FIO_RANDSEED * td->thread_number;
367
368	os_random_seed(seeds[4], &td->random_state);
369	return 0;
370}
371
372/*
373 * Adds a job to the list of things todo. Sanitizes the various options
374 * to make sure we don't have conflicts, and initializes various
375 * members of td.
376 */
377static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
378{
379	const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
380				   "randread", "randwrite", "randrw" };
381	unsigned int i;
382	const char *engine;
383	char fname[PATH_MAX];
384	int numjobs, file_alloced;
385
386	/*
387	 * the def_thread is just for options, it's not a real job
388	 */
389	if (td == &def_thread)
390		return 0;
391
392	/*
393	 * if we are just dumping the output command line, don't add the job
394	 */
395	if (dump_cmdline) {
396		put_job(td);
397		return 0;
398	}
399
400	engine = get_engine_name(td->o.ioengine);
401	td->io_ops = load_ioengine(td, engine);
402	if (!td->io_ops) {
403		log_err("fio: failed to load engine %s\n", engine);
404		goto err;
405	}
406
407	if (td->o.use_thread)
408		nr_thread++;
409	else
410		nr_process++;
411
412	if (td->o.odirect)
413		td->io_ops->flags |= FIO_RAWIO;
414
415	file_alloced = 0;
416	if (!td->o.filename && !td->files_index) {
417		file_alloced = 1;
418
419		if (td->o.nr_files == 1 && exists_and_not_file(jobname))
420			add_file(td, jobname);
421		else {
422			for (i = 0; i < td->o.nr_files; i++) {
423				sprintf(fname, "%s.%d.%d", jobname, td->thread_number, i);
424				add_file(td, fname);
425			}
426		}
427	}
428
429	if (fixup_options(td))
430		goto err;
431
432	if (td->io_ops->flags & FIO_DISKLESSIO) {
433		struct fio_file *f;
434
435		for_each_file(td, f, i)
436			f->real_file_size = -1ULL;
437	}
438
439	td->mutex = fio_sem_init(0);
440
441	td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
442	td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
443	td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
444	td->ddir_nr = td->o.ddir_nr;
445
446	if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
447	     && prev_group_jobs) {
448		prev_group_jobs = 0;
449		groupid++;
450	}
451
452	td->groupid = groupid;
453	prev_group_jobs++;
454
455	if (init_random_state(td))
456		goto err;
457
458	if (setup_rate(td))
459		goto err;
460
461	if (td->o.write_lat_log) {
462		setup_log(&td->ts.slat_log);
463		setup_log(&td->ts.clat_log);
464	}
465	if (td->o.write_bw_log)
466		setup_log(&td->ts.bw_log);
467
468	if (!td->o.name)
469		td->o.name = strdup(jobname);
470
471	if (!terse_output) {
472		if (!job_add_num) {
473			if (!strcmp(td->io_ops->name, "cpuio"))
474				log_info("%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->o.name, td->o.cpuload, td->o.cpucycle);
475			else {
476				char *c1, *c2, *c3, *c4;
477
478				c1 = to_kmg(td->o.min_bs[DDIR_READ]);
479				c2 = to_kmg(td->o.max_bs[DDIR_READ]);
480				c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
481				c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
482
483				log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->o.name, td->groupid, ddir_str[td->o.td_ddir], c1, c2, c3, c4, td->io_ops->name, td->o.iodepth);
484
485				free(c1);
486				free(c2);
487				free(c3);
488				free(c4);
489			}
490		} else if (job_add_num == 1)
491			log_info("...\n");
492	}
493
494	/*
495	 * recurse add identical jobs, clear numjobs and stonewall options
496	 * as they don't apply to sub-jobs
497	 */
498	numjobs = td->o.numjobs;
499	while (--numjobs) {
500		struct thread_data *td_new = get_new_job(0, td);
501
502		if (!td_new)
503			goto err;
504
505		td_new->o.numjobs = 1;
506		td_new->o.stonewall = 0;
507		td_new->o.new_group = 0;
508
509		if (file_alloced) {
510			td_new->o.filename = NULL;
511			td_new->files_index = 0;
512			td_new->files = NULL;
513		}
514
515		job_add_num = numjobs - 1;
516
517		if (add_job(td_new, jobname, job_add_num))
518			goto err;
519	}
520
521	return 0;
522err:
523	put_job(td);
524	return -1;
525}
526
527static int is_empty_or_comment(char *line)
528{
529	unsigned int i;
530
531	for (i = 0; i < strlen(line); i++) {
532		if (line[i] == ';')
533			return 1;
534		if (line[i] == '#')
535			return 1;
536		if (!isspace(line[i]) && !iscntrl(line[i]))
537			return 0;
538	}
539
540	return 1;
541}
542
543/*
544 * This is our [ini] type file parser.
545 */
546static int parse_jobs_ini(char *file, int stonewall_flag)
547{
548	unsigned int global;
549	struct thread_data *td;
550	char *string, *name;
551	fpos_t off;
552	FILE *f;
553	char *p;
554	int ret = 0, stonewall;
555	int first_sect = 1;
556
557	f = fopen(file, "r");
558	if (!f) {
559		perror("fopen job file");
560		return 1;
561	}
562
563	string = malloc(4096);
564
565	/*
566	 * it's really 256 + small bit, 280 should suffice
567	 */
568	name = malloc(280);
569	memset(name, 0, 280);
570
571	stonewall = stonewall_flag;
572	do {
573		p = fgets(string, 4095, f);
574		if (!p)
575			break;
576
577		strip_blank_front(&p);
578		strip_blank_end(p);
579
580		if (is_empty_or_comment(p))
581			continue;
582		if (sscanf(p, "[%255s]", name) != 1) {
583			log_err("fio: option <%s> outside of [] job section\n", p);
584			break;
585		}
586
587		global = !strncmp(name, "global", 6);
588
589		name[strlen(name) - 1] = '\0';
590
591		if (dump_cmdline) {
592			if (first_sect)
593				log_info("fio ");
594			if (!global)
595				log_info("--name=%s ", name);
596			first_sect = 0;
597		}
598
599		td = get_new_job(global, &def_thread);
600		if (!td) {
601			ret = 1;
602			break;
603		}
604
605		/*
606		 * Seperate multiple job files by a stonewall
607		 */
608		if (!global && stonewall) {
609			td->o.stonewall = stonewall;
610			stonewall = 0;
611		}
612
613		fgetpos(f, &off);
614		while ((p = fgets(string, 4096, f)) != NULL) {
615			if (is_empty_or_comment(p))
616				continue;
617
618			strip_blank_front(&p);
619
620			if (p[0] == '[')
621				break;
622
623			strip_blank_end(p);
624
625			fgetpos(f, &off);
626
627			/*
628			 * Don't break here, continue parsing options so we
629			 * dump all the bad ones. Makes trial/error fixups
630			 * easier on the user.
631			 */
632			ret |= fio_option_parse(td, p);
633			if (!ret && dump_cmdline)
634				log_info("--%s ", p);
635		}
636
637		if (!ret) {
638			fsetpos(f, &off);
639			ret = add_job(td, name, 0);
640		} else {
641			log_err("fio: job %s dropped\n", name);
642			put_job(td);
643		}
644	} while (!ret);
645
646	if (dump_cmdline)
647		log_info("\n");
648
649	free(string);
650	free(name);
651	fclose(f);
652	return ret;
653}
654
655static int fill_def_thread(void)
656{
657	memset(&def_thread, 0, sizeof(def_thread));
658
659	fio_getaffinity(getpid(), &def_thread.o.cpumask);
660
661	/*
662	 * fill default options
663	 */
664	fio_fill_default_options(&def_thread);
665
666	def_thread.o.timeout = def_timeout;
667	def_thread.o.write_bw_log = write_bw_log;
668	def_thread.o.write_lat_log = write_lat_log;
669
670	return 0;
671}
672
673static void free_shm(void)
674{
675	struct shmid_ds sbuf;
676
677	if (threads) {
678		shmdt((void *) threads);
679		threads = NULL;
680		shmctl(shm_id, IPC_RMID, &sbuf);
681	}
682}
683
684/*
685 * The thread area is shared between the main process and the job
686 * threads/processes. So setup a shared memory segment that will hold
687 * all the job info.
688 */
689static int setup_thread_area(void)
690{
691	/*
692	 * 1024 is too much on some machines, scale max_jobs if
693	 * we get a failure that looks like too large a shm segment
694	 */
695	do {
696		size_t size = max_jobs * sizeof(struct thread_data);
697
698		shm_id = shmget(0, size, IPC_CREAT | 0600);
699		if (shm_id != -1)
700			break;
701		if (errno != EINVAL) {
702			perror("shmget");
703			break;
704		}
705
706		max_jobs >>= 1;
707	} while (max_jobs);
708
709	if (shm_id == -1)
710		return 1;
711
712	threads = shmat(shm_id, NULL, 0);
713	if (threads == (void *) -1) {
714		perror("shmat");
715		return 1;
716	}
717
718	atexit(free_shm);
719	return 0;
720}
721
722static void usage(void)
723{
724	printf("%s\n", fio_version_string);
725	printf("\t--output\tWrite output to file\n");
726	printf("\t--timeout\tRuntime in seconds\n");
727	printf("\t--latency-log\tGenerate per-job latency logs\n");
728	printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
729	printf("\t--minimal\tMinimal (terse) output\n");
730	printf("\t--version\tPrint version info and exit\n");
731	printf("\t--help\t\tPrint this page\n");
732	printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
733	printf("\t--showcmd\tTurn a job file into command line options\n");
734}
735
736static int parse_cmd_line(int argc, char *argv[])
737{
738	struct thread_data *td = NULL;
739	int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
740
741	while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
742		switch (c) {
743		case 't':
744			def_timeout = atoi(optarg);
745			break;
746		case 'l':
747			write_lat_log = 1;
748			break;
749		case 'w':
750			write_bw_log = 1;
751			break;
752		case 'o':
753			f_out = fopen(optarg, "w+");
754			if (!f_out) {
755				perror("fopen output");
756				exit(1);
757			}
758			f_err = f_out;
759			break;
760		case 'm':
761			terse_output = 1;
762			break;
763		case 'h':
764			usage();
765			exit(0);
766		case 'c':
767			exit(fio_show_option_help(optarg));
768		case 's':
769			dump_cmdline = 1;
770			break;
771		case 'v':
772			printf("%s\n", fio_version_string);
773			exit(0);
774		case FIO_GETOPT_JOB: {
775			const char *opt = long_options[lidx].name;
776			char *val = optarg;
777
778			if (!strncmp(opt, "name", 4) && td) {
779				ret = add_job(td, td->o.name ?: "fio", 0);
780				if (ret) {
781					put_job(td);
782					return 0;
783				}
784				td = NULL;
785			}
786			if (!td) {
787				int global = 0;
788
789				if (strncmp(opt, "name", 4) ||
790				    !strncmp(val, "global", 6))
791					global = 1;
792
793				td = get_new_job(global, &def_thread);
794				if (!td)
795					return 0;
796			}
797
798			ret = fio_cmd_option_parse(td, opt, val);
799			if (ret)
800				dont_add_job = 1;
801			break;
802		}
803		default:
804			break;
805		}
806	}
807
808	if (td) {
809		if (dont_add_job)
810			put_job(td);
811		else {
812			ret = add_job(td, td->o.name ?: "fio", 0);
813			if (ret)
814				put_job(td);
815		}
816	}
817
818	while (optind < argc) {
819		ini_idx++;
820		ini_file = realloc(ini_file, ini_idx * sizeof(char *));
821		ini_file[ini_idx - 1] = strdup(argv[optind]);
822		optind++;
823	}
824
825	return ini_idx;
826}
827
828
829int parse_options(int argc, char *argv[])
830{
831	int job_files, i;
832
833	f_out = stdout;
834	f_err = stderr;
835
836	fio_options_dup_and_init(long_options);
837
838	if (setup_thread_area())
839		return 1;
840	if (fill_def_thread())
841		return 1;
842
843	job_files = parse_cmd_line(argc, argv);
844
845	for (i = 0; i < job_files; i++) {
846		if (fill_def_thread())
847			return 1;
848		if (parse_jobs_ini(ini_file[i], i))
849			return 1;
850		free(ini_file[i]);
851	}
852
853	free(ini_file);
854	options_mem_free(&def_thread);
855
856	if (!thread_number) {
857		if (dump_cmdline)
858			return 0;
859
860		log_err("No jobs defined(s)\n");
861		return 1;
862	}
863
864	return 0;
865}
866