init.c revision cf4464ca60ce75ac241c621bc70a2500beeb782f 
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 1.15.2";
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 (td_read(td) && !td_rw(td))
226		o->verify = 0;
227
228	if (o->norandommap && o->verify != VERIFY_NONE) {
229		log_err("fio: norandommap given, verify disabled\n");
230		o->verify = VERIFY_NONE;
231	}
232	if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
233		log_err("fio: bs_unaligned may not work with raw io\n");
234
235	/*
236	 * thinktime_spin must be less than thinktime
237	 */
238	if (o->thinktime_spin > o->thinktime)
239		o->thinktime_spin = o->thinktime;
240
241	/*
242	 * The low water mark cannot be bigger than the iodepth
243	 */
244	if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
245		/*
246		 * syslet work around - if the workload is sequential,
247		 * we want to let the queue drain all the way down to
248		 * avoid seeking between async threads
249		 */
250		if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
251			o->iodepth_low = 1;
252		else
253			o->iodepth_low = o->iodepth;
254	}
255
256	/*
257	 * If batch number isn't set, default to the same as iodepth
258	 */
259	if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
260		o->iodepth_batch = o->iodepth;
261
262	if (o->nr_files > td->files_index)
263		o->nr_files = td->files_index;
264
265	if (o->open_files > o->nr_files || !o->open_files)
266		o->open_files = o->nr_files;
267
268	if ((o->rate && o->rate_iops) || (o->ratemin && o->rate_iops_min)) {
269		log_err("fio: rate and rate_iops are mutually exclusive\n");
270		return 1;
271	}
272	if ((o->rate < o->ratemin) || (o->rate_iops < o->rate_iops_min)) {
273		log_err("fio: minimum rate exceeds rate\n");
274		return 1;
275	}
276
277	if (!o->timeout && o->time_based) {
278		log_err("fio: time_based requires a runtime/timeout setting\n");
279		o->time_based = 0;
280	}
281
282	return 0;
283}
284
285/*
286 * This function leaks the buffer
287 */
288static char *to_kmg(unsigned int val)
289{
290	char *buf = malloc(32);
291	char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
292	char *p = post;
293
294	do {
295		if (val & 1023)
296			break;
297
298		val >>= 10;
299		p++;
300	} while (*p);
301
302	snprintf(buf, 31, "%u%c", val, *p);
303	return buf;
304}
305
306/* External engines are specified by "external:name.o") */
307static const char *get_engine_name(const char *str)
308{
309	char *p = strstr(str, ":");
310
311	if (!p)
312		return str;
313
314	p++;
315	strip_blank_front(&p);
316	strip_blank_end(p);
317	return p;
318}
319
320static int exists_and_not_file(const char *filename)
321{
322	struct stat sb;
323
324	if (lstat(filename, &sb) == -1)
325		return 0;
326
327	if (S_ISREG(sb.st_mode))
328		return 0;
329
330	return 1;
331}
332
333/*
334 * Initialize the various random states we need (random io, block size ranges,
335 * read/write mix, etc).
336 */
337static int init_random_state(struct thread_data *td)
338{
339	unsigned long seeds[6];
340	int fd;
341
342	fd = open("/dev/urandom", O_RDONLY);
343	if (fd == -1) {
344		td_verror(td, errno, "open");
345		return 1;
346	}
347
348	if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
349		td_verror(td, EIO, "read");
350		close(fd);
351		return 1;
352	}
353
354	close(fd);
355
356	os_random_seed(seeds[0], &td->bsrange_state);
357	os_random_seed(seeds[1], &td->verify_state);
358	os_random_seed(seeds[2], &td->rwmix_state);
359
360	if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
361		os_random_seed(seeds[3], &td->next_file_state);
362
363	os_random_seed(seeds[5], &td->file_size_state);
364
365	if (!td_random(td))
366		return 0;
367
368	if (td->o.rand_repeatable)
369		seeds[4] = FIO_RANDSEED * td->thread_number;
370
371	os_random_seed(seeds[4], &td->random_state);
372	return 0;
373}
374
375/*
376 * Adds a job to the list of things todo. Sanitizes the various options
377 * to make sure we don't have conflicts, and initializes various
378 * members of td.
379 */
380static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
381{
382	const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
383				   "randread", "randwrite", "randrw" };
384	unsigned int i;
385	const char *engine;
386	char fname[PATH_MAX];
387	int numjobs, file_alloced;
388
389	/*
390	 * the def_thread is just for options, it's not a real job
391	 */
392	if (td == &def_thread)
393		return 0;
394
395	/*
396	 * if we are just dumping the output command line, don't add the job
397	 */
398	if (dump_cmdline) {
399		put_job(td);
400		return 0;
401	}
402
403	engine = get_engine_name(td->o.ioengine);
404	td->io_ops = load_ioengine(td, engine);
405	if (!td->io_ops) {
406		log_err("fio: failed to load engine %s\n", engine);
407		goto err;
408	}
409
410	if (td->o.use_thread)
411		nr_thread++;
412	else
413		nr_process++;
414
415	if (td->o.odirect)
416		td->io_ops->flags |= FIO_RAWIO;
417
418	file_alloced = 0;
419	if (!td->o.filename && !td->files_index) {
420		file_alloced = 1;
421
422		if (td->o.nr_files == 1 && exists_and_not_file(jobname))
423			add_file(td, jobname);
424		else {
425			for (i = 0; i < td->o.nr_files; i++) {
426				sprintf(fname, "%s.%d.%d", jobname, td->thread_number, i);
427				add_file(td, fname);
428			}
429		}
430	}
431
432	if (fixup_options(td))
433		goto err;
434
435	if (td->io_ops->flags & FIO_DISKLESSIO) {
436		struct fio_file *f;
437
438		for_each_file(td, f, i)
439			f->real_file_size = -1ULL;
440	}
441
442	td->mutex = fio_sem_init(0);
443
444	td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
445	td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
446	td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
447	td->ddir_nr = td->o.ddir_nr;
448
449	if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
450	     && prev_group_jobs) {
451		prev_group_jobs = 0;
452		groupid++;
453	}
454
455	td->groupid = groupid;
456	prev_group_jobs++;
457
458	if (init_random_state(td))
459		goto err;
460
461	if (setup_rate(td))
462		goto err;
463
464	if (td->o.write_lat_log) {
465		setup_log(&td->ts.slat_log);
466		setup_log(&td->ts.clat_log);
467	}
468	if (td->o.write_bw_log)
469		setup_log(&td->ts.bw_log);
470
471	if (!td->o.name)
472		td->o.name = strdup(jobname);
473
474	if (!terse_output) {
475		if (!job_add_num) {
476			if (!strcmp(td->io_ops->name, "cpuio"))
477				log_info("%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->o.name, td->o.cpuload, td->o.cpucycle);
478			else {
479				char *c1, *c2, *c3, *c4;
480
481				c1 = to_kmg(td->o.min_bs[DDIR_READ]);
482				c2 = to_kmg(td->o.max_bs[DDIR_READ]);
483				c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
484				c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
485
486				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);
487
488				free(c1);
489				free(c2);
490				free(c3);
491				free(c4);
492			}
493		} else if (job_add_num == 1)
494			log_info("...\n");
495	}
496
497	/*
498	 * recurse add identical jobs, clear numjobs and stonewall options
499	 * as they don't apply to sub-jobs
500	 */
501	numjobs = td->o.numjobs;
502	while (--numjobs) {
503		struct thread_data *td_new = get_new_job(0, td);
504
505		if (!td_new)
506			goto err;
507
508		td_new->o.numjobs = 1;
509		td_new->o.stonewall = 0;
510		td_new->o.new_group = 0;
511
512		if (file_alloced) {
513			td_new->o.filename = NULL;
514			td_new->files_index = 0;
515			td_new->files = NULL;
516		}
517
518		job_add_num = numjobs - 1;
519
520		if (add_job(td_new, jobname, job_add_num))
521			goto err;
522	}
523
524	return 0;
525err:
526	put_job(td);
527	return -1;
528}
529
530static int is_empty_or_comment(char *line)
531{
532	unsigned int i;
533
534	for (i = 0; i < strlen(line); i++) {
535		if (line[i] == ';')
536			return 1;
537		if (line[i] == '#')
538			return 1;
539		if (!isspace(line[i]) && !iscntrl(line[i]))
540			return 0;
541	}
542
543	return 1;
544}
545
546/*
547 * This is our [ini] type file parser.
548 */
549static int parse_jobs_ini(char *file, int stonewall_flag)
550{
551	unsigned int global;
552	struct thread_data *td;
553	char *string, *name;
554	fpos_t off;
555	FILE *f;
556	char *p;
557	int ret = 0, stonewall;
558
559	f = fopen(file, "r");
560	if (!f) {
561		perror("fopen job file");
562		return 1;
563	}
564
565	string = malloc(4096);
566	name = malloc(256);
567	memset(name, 0, 256);
568
569	stonewall = stonewall_flag;
570	do {
571		p = fgets(string, 4095, f);
572		if (!p)
573			break;
574
575		strip_blank_front(&p);
576		strip_blank_end(p);
577
578		if (is_empty_or_comment(p))
579			continue;
580		if (sscanf(p, "[%255s]", name) != 1) {
581			log_err("fio: option <%s> outside of job section\n", p);
582			continue;
583		}
584
585		global = !strncmp(name, "global", 6);
586
587		name[strlen(name) - 1] = '\0';
588
589		if (dump_cmdline) {
590			log_info("fio ");
591			if (!global)
592				log_info("--name=%s ", name);
593		}
594
595		td = get_new_job(global, &def_thread);
596		if (!td) {
597			ret = 1;
598			break;
599		}
600
601		/*
602		 * Seperate multiple job files by a stonewall
603		 */
604		if (!global && stonewall) {
605			td->o.stonewall = stonewall;
606			stonewall = 0;
607		}
608
609		fgetpos(f, &off);
610		while ((p = fgets(string, 4096, f)) != NULL) {
611			if (is_empty_or_comment(p))
612				continue;
613
614			strip_blank_front(&p);
615
616			if (p[0] == '[')
617				break;
618
619			strip_blank_end(p);
620
621			fgetpos(f, &off);
622
623			/*
624			 * Don't break here, continue parsing options so we
625			 * dump all the bad ones. Makes trial/error fixups
626			 * easier on the user.
627			 */
628			ret |= fio_option_parse(td, p);
629			if (!ret && dump_cmdline)
630				log_info("--%s ", p);
631		}
632
633		if (!ret) {
634			fsetpos(f, &off);
635			ret = add_job(td, name, 0);
636		} else {
637			log_err("fio: job %s dropped\n", name);
638			put_job(td);
639		}
640	} while (!ret);
641
642	if (dump_cmdline)
643		log_info("\n");
644
645	free(string);
646	free(name);
647	fclose(f);
648	return ret;
649}
650
651static int fill_def_thread(void)
652{
653	memset(&def_thread, 0, sizeof(def_thread));
654
655	if (fio_getaffinity(getpid(), &def_thread.o.cpumask) == -1) {
656		perror("sched_getaffinity");
657		return 1;
658	}
659
660	/*
661	 * fill default options
662	 */
663	fio_fill_default_options(&def_thread);
664
665	def_thread.o.timeout = def_timeout;
666	def_thread.o.write_bw_log = write_bw_log;
667	def_thread.o.write_lat_log = write_lat_log;
668
669#ifdef FIO_HAVE_DISK_UTIL
670	def_thread.o.do_disk_util = 1;
671#endif
672
673	return 0;
674}
675
676static void free_shm(void)
677{
678	struct shmid_ds sbuf;
679
680	if (threads) {
681		shmdt((void *) threads);
682		threads = NULL;
683		shmctl(shm_id, IPC_RMID, &sbuf);
684	}
685}
686
687/*
688 * The thread area is shared between the main process and the job
689 * threads/processes. So setup a shared memory segment that will hold
690 * all the job info.
691 */
692static int setup_thread_area(void)
693{
694	/*
695	 * 1024 is too much on some machines, scale max_jobs if
696	 * we get a failure that looks like too large a shm segment
697	 */
698	do {
699		size_t size = max_jobs * sizeof(struct thread_data);
700
701		shm_id = shmget(0, size, IPC_CREAT | 0600);
702		if (shm_id != -1)
703			break;
704		if (errno != EINVAL) {
705			perror("shmget");
706			break;
707		}
708
709		max_jobs >>= 1;
710	} while (max_jobs);
711
712	if (shm_id == -1)
713		return 1;
714
715	threads = shmat(shm_id, NULL, 0);
716	if (threads == (void *) -1) {
717		perror("shmat");
718		return 1;
719	}
720
721	atexit(free_shm);
722	return 0;
723}
724
725static void usage(void)
726{
727	printf("%s\n", fio_version_string);
728	printf("\t--output\tWrite output to file\n");
729	printf("\t--timeout\tRuntime in seconds\n");
730	printf("\t--latency-log\tGenerate per-job latency logs\n");
731	printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
732	printf("\t--minimal\tMinimal (terse) output\n");
733	printf("\t--version\tPrint version info and exit\n");
734	printf("\t--help\t\tPrint this page\n");
735	printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
736	printf("\t--showcmd\tTurn a job file into command line options\n");
737}
738
739static int parse_cmd_line(int argc, char *argv[])
740{
741	struct thread_data *td = NULL;
742	int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
743
744	while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
745		switch (c) {
746		case 't':
747			def_timeout = atoi(optarg);
748			break;
749		case 'l':
750			write_lat_log = 1;
751			break;
752		case 'w':
753			write_bw_log = 1;
754			break;
755		case 'o':
756			f_out = fopen(optarg, "w+");
757			if (!f_out) {
758				perror("fopen output");
759				exit(1);
760			}
761			f_err = f_out;
762			break;
763		case 'm':
764			terse_output = 1;
765			break;
766		case 'h':
767			usage();
768			exit(0);
769		case 'c':
770			exit(fio_show_option_help(optarg));
771		case 's':
772			dump_cmdline = 1;
773			break;
774		case 'v':
775			printf("%s\n", fio_version_string);
776			exit(0);
777		case FIO_GETOPT_JOB: {
778			const char *opt = long_options[lidx].name;
779			char *val = optarg;
780
781			if (!strncmp(opt, "name", 4) && td) {
782				ret = add_job(td, td->o.name ?: "fio", 0);
783				if (ret) {
784					put_job(td);
785					return 0;
786				}
787				td = NULL;
788			}
789			if (!td) {
790				int global = !strncmp(val, "global", 6);
791
792				td = get_new_job(global, &def_thread);
793				if (!td)
794					return 0;
795			}
796
797			ret = fio_cmd_option_parse(td, opt, val);
798			if (ret)
799				dont_add_job = 1;
800			break;
801		}
802		default:
803			break;
804		}
805	}
806
807	if (td) {
808		if (dont_add_job)
809			put_job(td);
810		else {
811			ret = add_job(td, td->o.name ?: "fio", 0);
812			if (ret)
813				put_job(td);
814		}
815	}
816
817	while (optind < argc) {
818		ini_idx++;
819		ini_file = realloc(ini_file, ini_idx * sizeof(char *));
820		ini_file[ini_idx - 1] = strdup(argv[optind]);
821		optind++;
822	}
823
824	return ini_idx;
825}
826
827
828int parse_options(int argc, char *argv[])
829{
830	int job_files, i;
831
832	f_out = stdout;
833	f_err = stderr;
834
835	fio_options_dup_and_init(long_options);
836
837	if (setup_thread_area())
838		return 1;
839	if (fill_def_thread())
840		return 1;
841
842	job_files = parse_cmd_line(argc, argv);
843
844	for (i = 0; i < job_files; i++) {
845		if (fill_def_thread())
846			return 1;
847		if (parse_jobs_ini(ini_file[i], i))
848			return 1;
849		free(ini_file[i]);
850	}
851
852	free(ini_file);
853	options_mem_free(&def_thread);
854
855	if (!thread_number) {
856		if (dump_cmdline)
857			return 0;
858
859		log_err("No jobs defined(s)\n");
860		return 1;
861	}
862
863	return 0;
864}
865