init.c revision 61697c37566a211ad3fd70f46397d9b7675b0fc2 
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#define FIO_RANDSEED		(0xb1899bedUL)
22
23#define td_var_offset(var)	((size_t) &((struct thread_data *)0)->var)
24
25static int str_rw_cb(void *, const char *);
26static int str_ioengine_cb(void *, const char *);
27static int str_mem_cb(void *, const char *);
28static int str_verify_cb(void *, const char *);
29static int str_lockmem_cb(void *, unsigned long *);
30#ifdef FIO_HAVE_IOPRIO
31static int str_prio_cb(void *, unsigned int *);
32static int str_prioclass_cb(void *, unsigned int *);
33#endif
34static int str_exitall_cb(void);
35static int str_cpumask_cb(void *, unsigned int *);
36
37#define __stringify_1(x)	#x
38#define __stringify(x)		__stringify_1(x)
39
40/*
41 * Map of job/command line options
42 */
43static struct fio_option options[] = {
44	{
45		.name	= "description",
46		.type	= FIO_OPT_STR_STORE,
47		.off1	= td_var_offset(description),
48		.help	= "Text job description",
49	},
50	{
51		.name	= "name",
52		.type	= FIO_OPT_STR_STORE,
53		.off1	= td_var_offset(name),
54		.help	= "Name of this job",
55	},
56	{
57		.name	= "directory",
58		.type	= FIO_OPT_STR_STORE,
59		.off1	= td_var_offset(directory),
60		.help	= "Directory to store files in",
61	},
62	{
63		.name	= "filename",
64		.type	= FIO_OPT_STR_STORE,
65		.off1	= td_var_offset(filename),
66		.help	= "Force the use of a specific file",
67	},
68	{
69		.name	= "rw",
70		.type	= FIO_OPT_STR,
71		.cb	= str_rw_cb,
72		.help	= "IO direction",
73		.def	= "read",
74		.posval	= { "read", "write", "randwrite", "randread", "rw",
75				"randrw", },
76	},
77	{
78		.name	= "ioengine",
79		.type	= FIO_OPT_STR,
80		.cb	= str_ioengine_cb,
81		.help	= "IO engine to use",
82		.def	= "sync",
83		.posval	= { "sync", "libaio", "posixaio", "mmap", "splice",
84				"sg", "null", },
85	},
86	{
87		.name	= "iodepth",
88		.type	= FIO_OPT_INT,
89		.off1	= td_var_offset(iodepth),
90		.help	= "Amount of IO buffers to keep in flight",
91		.def	= "1",
92	},
93	{
94		.name	= "size",
95		.type	= FIO_OPT_STR_VAL,
96		.off1	= td_var_offset(total_file_size),
97		.help	= "Size of device or file",
98	},
99	{
100		.name	= "bs",
101		.type	= FIO_OPT_STR_VAL_INT,
102		.off1	= td_var_offset(bs[DDIR_READ]),
103		.off2	= td_var_offset(bs[DDIR_WRITE]),
104		.help	= "Block size unit",
105		.def	= "4k",
106	},
107	{
108		.name	= "bsrange",
109		.type	= FIO_OPT_RANGE,
110		.off1	= td_var_offset(min_bs[DDIR_READ]),
111		.off2	= td_var_offset(max_bs[DDIR_READ]),
112		.off3	= td_var_offset(min_bs[DDIR_WRITE]),
113		.off4	= td_var_offset(max_bs[DDIR_WRITE]),
114		.help	= "Set block size range (in more detail than bs)",
115	},
116	{
117		.name	= "bs_unaligned",
118		.type	= FIO_OPT_STR_SET,
119		.off1	= td_var_offset(bs_unaligned),
120		.help	= "Don't sector align IO buffer sizes",
121	},
122	{
123		.name	= "offset",
124		.type	= FIO_OPT_STR_VAL,
125		.off1	= td_var_offset(start_offset),
126		.help	= "Start IO from this offset",
127		.def	= "0",
128	},
129	{
130		.name	= "randrepeat",
131		.type	= FIO_OPT_BOOL,
132		.off1	= td_var_offset(rand_repeatable),
133		.help	= "Use repeatable random IO pattern",
134		.def	= "1",
135	},
136	{
137		.name	= "norandommap",
138		.type	= FIO_OPT_STR_SET,
139		.off1	= td_var_offset(norandommap),
140		.help	= "Accept potential duplicate random blocks",
141	},
142	{
143		.name	= "nrfiles",
144		.type	= FIO_OPT_INT,
145		.off1	= td_var_offset(nr_files),
146		.help	= "Split job workload between this number of files",
147		.def	= "1",
148	},
149	{
150		.name	= "fsync",
151		.type	= FIO_OPT_INT,
152		.off1	= td_var_offset(fsync_blocks),
153		.help	= "Issue fsync for writes every given number of blocks",
154		.def	= "0",
155	},
156	{
157		.name	= "direct",
158		.type	= FIO_OPT_BOOL,
159		.off1	= td_var_offset(odirect),
160		.help	= "Use O_DIRECT IO (negates buffered)",
161		.def	= "0",
162	},
163	{
164		.name	= "buffered",
165		.type	= FIO_OPT_BOOL,
166		.off1	= td_var_offset(odirect),
167		.neg	= 1,
168		.help	= "Use buffered IO (negates direct)",
169		.def	= "1",
170	},
171	{
172		.name	= "overwrite",
173		.type	= FIO_OPT_BOOL,
174		.off1	= td_var_offset(overwrite),
175		.help	= "When writing, set whether to overwrite current data",
176		.def	= "0",
177	},
178	{
179		.name	= "loops",
180		.type	= FIO_OPT_INT,
181		.off1	= td_var_offset(loops),
182		.help	= "Number of times to run the job",
183		.def	= "1",
184	},
185	{
186		.name	= "numjobs",
187		.type	= FIO_OPT_INT,
188		.off1	= td_var_offset(numjobs),
189		.help	= "Duplicate this job this many times",
190		.def	= "1",
191	},
192	{
193		.name	= "startdelay",
194		.type	= FIO_OPT_INT,
195		.off1	= td_var_offset(start_delay),
196		.help	= "Only start job when this period has passed",
197		.def	= "0",
198	},
199	{
200		.name	= "runtime",
201		.alias	= "timeout",
202		.type	= FIO_OPT_STR_VAL_TIME,
203		.off1	= td_var_offset(timeout),
204		.help	= "Stop workload when this amount of time has passed",
205		.def	= "0",
206	},
207	{
208		.name	= "mem",
209		.type	= FIO_OPT_STR,
210		.cb	= str_mem_cb,
211		.help	= "Backing type for IO buffers",
212		.def	= "malloc",
213		.posval	=  { "malloc", "shm", "shmhuge", "mmap", "mmaphuge", },
214	},
215	{
216		.name	= "verify",
217		.type	= FIO_OPT_STR,
218		.cb	= str_verify_cb,
219		.help	= "Verify sum function",
220		.def	= "0",
221		.posval	= { "crc32", "md5", },
222	},
223	{
224		.name	= "write_iolog",
225		.type	= FIO_OPT_STR_STORE,
226		.off1	= td_var_offset(write_iolog_file),
227		.help	= "Store IO pattern to file",
228	},
229	{
230		.name	= "read_iolog",
231		.type	= FIO_OPT_STR_STORE,
232		.off1	= td_var_offset(read_iolog_file),
233		.help	= "Playback IO pattern from file",
234	},
235	{
236		.name	= "exec_prerun",
237		.type	= FIO_OPT_STR_STORE,
238		.off1	= td_var_offset(exec_prerun),
239		.help	= "Execute this file prior to running job",
240	},
241	{
242		.name	= "exec_postrun",
243		.type	= FIO_OPT_STR_STORE,
244		.off1	= td_var_offset(exec_postrun),
245		.help	= "Execute this file after running job",
246	},
247#ifdef FIO_HAVE_IOSCHED_SWITCH
248	{
249		.name	= "ioscheduler",
250		.type	= FIO_OPT_STR_STORE,
251		.off1	= td_var_offset(ioscheduler),
252		.help	= "Use this IO scheduler on the backing device",
253	},
254#endif
255	{
256		.name	= "zonesize",
257		.type	= FIO_OPT_STR_VAL,
258		.off1	= td_var_offset(zone_size),
259		.help	= "Give size of an IO zone",
260		.def	= "0",
261	},
262	{
263		.name	= "zoneskip",
264		.type	= FIO_OPT_STR_VAL,
265		.off1	= td_var_offset(zone_skip),
266		.help	= "Space between IO zones",
267		.def	= "0",
268	},
269	{
270		.name	= "lockmem",
271		.type	= FIO_OPT_STR_VAL,
272		.cb	= str_lockmem_cb,
273		.help	= "Lock down this amount of memory",
274		.def	= "0",
275	},
276	{
277		.name	= "rwmixcycle",
278		.type	= FIO_OPT_INT,
279		.off1	= td_var_offset(rwmixcycle),
280		.help	= "Cycle period for mixed read/write workloads (msec)",
281		.def	= "500",
282	},
283	{
284		.name	= "rwmixread",
285		.type	= FIO_OPT_INT,
286		.off1	= td_var_offset(rwmixread),
287		.maxval	= 100,
288		.help	= "Percentage of mixed workload that is reads",
289		.def	= "50",
290	},
291	{
292		.name	= "rwmixwrite",
293		.type	= FIO_OPT_INT,
294		.off1	= td_var_offset(rwmixwrite),
295		.maxval	= 100,
296		.help	= "Percentage of mixed workload that is writes",
297		.def	= "50",
298	},
299	{
300		.name	= "nice",
301		.type	= FIO_OPT_INT,
302		.off1	= td_var_offset(nice),
303		.help	= "Set job CPU nice value",
304		.minval	= -19,
305		.maxval	= 20,
306		.def	= "0",
307	},
308#ifdef FIO_HAVE_IOPRIO
309	{
310		.name	= "prio",
311		.type	= FIO_OPT_INT,
312		.cb	= str_prio_cb,
313		.help	= "Set job IO priority value",
314		.minval	= 0,
315		.maxval	= 7,
316	},
317	{
318		.name	= "prioclass",
319		.type	= FIO_OPT_INT,
320		.cb	= str_prioclass_cb,
321		.help	= "Set job IO priority class",
322		.minval	= 0,
323		.maxval	= 3,
324	},
325#endif
326	{
327		.name	= "thinktime",
328		.type	= FIO_OPT_INT,
329		.off1	= td_var_offset(thinktime),
330		.help	= "Idle time between IO buffers (usec)",
331		.def	= "0",
332	},
333	{
334		.name	= "thinktime_blocks",
335		.type	= FIO_OPT_INT,
336		.off1	= td_var_offset(thinktime_blocks),
337		.help	= "IO buffer period between 'thinktime'",
338		.def	= "1",
339	},
340	{
341		.name	= "rate",
342		.type	= FIO_OPT_INT,
343		.off1	= td_var_offset(rate),
344		.help	= "Set bandwidth rate",
345	},
346	{
347		.name	= "ratemin",
348		.type	= FIO_OPT_INT,
349		.off1	= td_var_offset(ratemin),
350		.help	= "The bottom limit accepted",
351	},
352	{
353		.name	= "ratecycle",
354		.type	= FIO_OPT_INT,
355		.off1	= td_var_offset(ratecycle),
356		.help	= "Window average for rate limits (msec)",
357		.def	= "1000",
358	},
359	{
360		.name	= "invalidate",
361		.type	= FIO_OPT_BOOL,
362		.off1	= td_var_offset(invalidate_cache),
363		.help	= "Invalidate buffer/page cache prior to running job",
364		.def	= "1",
365	},
366	{
367		.name	= "sync",
368		.type	= FIO_OPT_BOOL,
369		.off1	= td_var_offset(sync_io),
370		.help	= "Use O_SYNC for buffered writes",
371		.def	= "0",
372	},
373	{
374		.name	= "bwavgtime",
375		.type	= FIO_OPT_INT,
376		.off1	= td_var_offset(bw_avg_time),
377		.help	= "Time window over which to calculate bandwidth (msec)",
378		.def	= "500",
379	},
380	{
381		.name	= "create_serialize",
382		.type	= FIO_OPT_BOOL,
383		.off1	= td_var_offset(create_serialize),
384		.help	= "Serialize creating of job files",
385		.def	= "1",
386	},
387	{
388		.name	= "create_fsync",
389		.type	= FIO_OPT_BOOL,
390		.off1	= td_var_offset(create_fsync),
391		.help	= "Fsync file after creation",
392		.def	= "1",
393	},
394	{
395		.name	= "cpuload",
396		.type	= FIO_OPT_INT,
397		.off1	= td_var_offset(cpuload),
398		.help	= "Use this percentage of CPU",
399	},
400	{
401		.name	= "cpuchunks",
402		.type	= FIO_OPT_INT,
403		.off1	= td_var_offset(cpucycle),
404		.help	= "Length of the CPU burn cycles",
405	},
406#ifdef FIO_HAVE_CPU_AFFINITY
407	{
408		.name	= "cpumask",
409		.type	= FIO_OPT_INT,
410		.cb	= str_cpumask_cb,
411		.help	= "CPU affinity mask",
412	},
413#endif
414	{
415		.name	= "end_fsync",
416		.type	= FIO_OPT_BOOL,
417		.off1	= td_var_offset(end_fsync),
418		.help	= "Include fsync at the end of job",
419		.def	= "0",
420	},
421	{
422		.name	= "unlink",
423		.type	= FIO_OPT_BOOL,
424		.off1	= td_var_offset(unlink),
425		.help	= "Unlink created files after job has completed",
426		.def	= "0",
427	},
428	{
429		.name	= "exitall",
430		.type	= FIO_OPT_STR_SET,
431		.cb	= str_exitall_cb,
432		.help	= "Terminate all jobs when one exits",
433	},
434	{
435		.name	= "stonewall",
436		.type	= FIO_OPT_STR_SET,
437		.off1	= td_var_offset(stonewall),
438		.help	= "Insert a hard barrier between this job and previous",
439	},
440	{
441		.name	= "thread",
442		.type	= FIO_OPT_STR_SET,
443		.off1	= td_var_offset(use_thread),
444		.help	= "Use threads instead of forks",
445	},
446	{
447		.name	= "write_bw_log",
448		.type	= FIO_OPT_STR_SET,
449		.off1	= td_var_offset(write_bw_log),
450		.help	= "Write log of bandwidth during run",
451	},
452	{
453		.name	= "write_lat_log",
454		.type	= FIO_OPT_STR_SET,
455		.off1	= td_var_offset(write_lat_log),
456		.help	= "Write log of latency during run",
457	},
458	{
459		.name	= "hugepage-size",
460		.type	= FIO_OPT_STR_VAL,
461		.off1	= td_var_offset(hugepage_size),
462		.help	= "When using hugepages, specify size of each page",
463		.def	= __stringify(FIO_HUGE_PAGE),
464	},
465	{
466		.name = NULL,
467	},
468};
469
470#define FIO_JOB_OPTS	(sizeof(options) / sizeof(struct fio_option))
471#define FIO_CMD_OPTS	(16)
472#define FIO_GETOPT_JOB	(0x89988998)
473
474/*
475 * Command line options. These will contain the above, plus a few
476 * extra that only pertain to fio itself and not jobs.
477 */
478static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
479	{
480		.name		= "output",
481		.has_arg	= required_argument,
482		.val		= 'o',
483	},
484	{
485		.name		= "timeout",
486		.has_arg	= required_argument,
487		.val		= 't',
488	},
489	{
490		.name		= "latency-log",
491		.has_arg	= required_argument,
492		.val		= 'l',
493	},
494	{
495		.name		= "bandwidth-log",
496		.has_arg	= required_argument,
497		.val		= 'b',
498	},
499	{
500		.name		= "minimal",
501		.has_arg	= optional_argument,
502		.val		= 'm',
503	},
504	{
505		.name		= "version",
506		.has_arg	= no_argument,
507		.val		= 'v',
508	},
509	{
510		.name		= "help",
511		.has_arg	= no_argument,
512		.val		= 'h',
513	},
514	{
515		.name		= "cmdhelp",
516		.has_arg	= required_argument,
517		.val		= 'c',
518	},
519	{
520		.name		= NULL,
521	},
522};
523
524static int def_timeout = 0;
525
526static char fio_version_string[] = "fio 1.11";
527
528static char **ini_file;
529static int max_jobs = MAX_JOBS;
530
531struct thread_data def_thread;
532struct thread_data *threads = NULL;
533
534int exitall_on_terminate = 0;
535int terse_output = 0;
536unsigned long long mlock_size = 0;
537FILE *f_out = NULL;
538FILE *f_err = NULL;
539
540static int write_lat_log = 0;
541int write_bw_log = 0;
542
543/*
544 * Return a free job structure.
545 */
546static struct thread_data *get_new_job(int global, struct thread_data *parent)
547{
548	struct thread_data *td;
549
550	if (global)
551		return &def_thread;
552	if (thread_number >= max_jobs)
553		return NULL;
554
555	td = &threads[thread_number++];
556	*td = *parent;
557
558	td->thread_number = thread_number;
559	return td;
560}
561
562static void put_job(struct thread_data *td)
563{
564	if (td == &def_thread)
565		return;
566
567	memset(&threads[td->thread_number - 1], 0, sizeof(*td));
568	thread_number--;
569}
570
571/*
572 * Lazy way of fixing up options that depend on each other. We could also
573 * define option callback handlers, but this is easier.
574 */
575static void fixup_options(struct thread_data *td)
576{
577	if (!td->rwmixread && td->rwmixwrite)
578		td->rwmixread = 100 - td->rwmixwrite;
579
580	if (td->write_iolog_file && td->read_iolog_file) {
581		log_err("fio: read iolog overrides write_iolog\n");
582		free(td->write_iolog_file);
583		td->write_iolog_file = NULL;
584	}
585
586	if (td->io_ops->flags & FIO_SYNCIO)
587		td->iodepth = 1;
588	else {
589		if (!td->iodepth)
590			td->iodepth = td->nr_files;
591	}
592
593	/*
594	 * only really works for sequential io for now, and with 1 file
595	 */
596	if (td->zone_size && !td->sequential && td->nr_files == 1)
597		td->zone_size = 0;
598
599	/*
600	 * Reads can do overwrites, we always need to pre-create the file
601	 */
602	if (td_read(td) || td_rw(td))
603		td->overwrite = 1;
604
605	if (!td->min_bs[DDIR_READ])
606		td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
607	if (!td->max_bs[DDIR_READ])
608		td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
609	if (!td->min_bs[DDIR_WRITE])
610		td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
611	if (!td->max_bs[DDIR_WRITE])
612		td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
613
614	td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
615
616	if (td_read(td) && !td_rw(td))
617		td->verify = 0;
618
619	if (td->norandommap && td->verify != VERIFY_NONE) {
620		log_err("fio: norandommap given, verify disabled\n");
621		td->verify = VERIFY_NONE;
622	}
623	if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
624		log_err("fio: bs_unaligned may not work with raw io\n");
625
626	/*
627	 * O_DIRECT and char doesn't mix, clear that flag if necessary.
628	 */
629	if (td->filetype == FIO_TYPE_CHAR && td->odirect)
630		td->odirect = 0;
631}
632
633/*
634 * This function leaks the buffer
635 */
636static char *to_kmg(unsigned int val)
637{
638	char *buf = malloc(32);
639	char post[] = { 0, 'K', 'M', 'G', 'P', 0 };
640	char *p = post;
641
642	do {
643		if (val & 1023)
644			break;
645
646		val >>= 10;
647		p++;
648	} while (*p);
649
650	snprintf(buf, 31, "%u%c", val, *p);
651	return buf;
652}
653
654/*
655 * Adds a job to the list of things todo. Sanitizes the various options
656 * to make sure we don't have conflicts, and initializes various
657 * members of td.
658 */
659static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
660{
661	const char *ddir_str[] = { "read", "write", "randread", "randwrite",
662				   "rw", NULL, "randrw" };
663	struct stat sb;
664	int numjobs, ddir, i;
665	struct fio_file *f;
666
667	/*
668	 * the def_thread is just for options, it's not a real job
669	 */
670	if (td == &def_thread)
671		return 0;
672
673	assert(td->io_ops);
674
675	if (td->odirect)
676		td->io_ops->flags |= FIO_RAWIO;
677
678	td->filetype = FIO_TYPE_FILE;
679	if (!stat(jobname, &sb)) {
680		if (S_ISBLK(sb.st_mode))
681			td->filetype = FIO_TYPE_BD;
682		else if (S_ISCHR(sb.st_mode))
683			td->filetype = FIO_TYPE_CHAR;
684	}
685
686	fixup_options(td);
687
688	if (td->filename)
689		td->nr_uniq_files = 1;
690	else
691		td->nr_uniq_files = td->nr_files;
692
693	if (td->filetype == FIO_TYPE_FILE || td->filename) {
694		char tmp[PATH_MAX];
695		int len = 0;
696
697		if (td->directory && td->directory[0] != '\0')
698			len = sprintf(tmp, "%s/", td->directory);
699
700		td->files = malloc(sizeof(struct fio_file) * td->nr_files);
701
702		for_each_file(td, f, i) {
703			memset(f, 0, sizeof(*f));
704			f->fd = -1;
705
706			if (td->filename)
707				sprintf(tmp + len, "%s", td->filename);
708			else
709				sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
710			f->file_name = strdup(tmp);
711		}
712	} else {
713		td->nr_files = 1;
714		td->files = malloc(sizeof(struct fio_file));
715		f = &td->files[0];
716
717		memset(f, 0, sizeof(*f));
718		f->fd = -1;
719		f->file_name = strdup(jobname);
720	}
721
722	for_each_file(td, f, i) {
723		f->file_size = td->total_file_size / td->nr_files;
724		f->file_offset = td->start_offset;
725	}
726
727	fio_sem_init(&td->mutex, 0);
728
729	td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
730	td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
731	td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
732
733	if (td->stonewall && td->thread_number > 1)
734		groupid++;
735
736	td->groupid = groupid;
737
738	if (setup_rate(td))
739		goto err;
740
741	if (td->write_lat_log) {
742		setup_log(&td->slat_log);
743		setup_log(&td->clat_log);
744	}
745	if (td->write_bw_log)
746		setup_log(&td->bw_log);
747
748	if (!td->name)
749		td->name = strdup(jobname);
750
751	ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
752
753	if (!terse_output) {
754		if (!job_add_num) {
755			if (td->io_ops->flags & FIO_CPUIO)
756				fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
757			else {
758				char *c1, *c2, *c3, *c4;
759
760				c1 = to_kmg(td->min_bs[DDIR_READ]);
761				c2 = to_kmg(td->max_bs[DDIR_READ]);
762				c3 = to_kmg(td->min_bs[DDIR_WRITE]);
763				c4 = to_kmg(td->max_bs[DDIR_WRITE]);
764
765				fprintf(f_out, "%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[ddir], c1, c2, c3, c4, td->io_ops->name, td->iodepth);
766
767				free(c1);
768				free(c2);
769				free(c3);
770				free(c4);
771			}
772		} else if (job_add_num == 1)
773			fprintf(f_out, "...\n");
774	}
775
776	/*
777	 * recurse add identical jobs, clear numjobs and stonewall options
778	 * as they don't apply to sub-jobs
779	 */
780	numjobs = td->numjobs;
781	while (--numjobs) {
782		struct thread_data *td_new = get_new_job(0, td);
783
784		if (!td_new)
785			goto err;
786
787		td_new->numjobs = 1;
788		td_new->stonewall = 0;
789		job_add_num = numjobs - 1;
790
791		if (add_job(td_new, jobname, job_add_num))
792			goto err;
793	}
794	return 0;
795err:
796	put_job(td);
797	return -1;
798}
799
800/*
801 * Initialize the various random states we need (random io, block size ranges,
802 * read/write mix, etc).
803 */
804int init_random_state(struct thread_data *td)
805{
806	unsigned long seeds[4];
807	int fd, num_maps, blocks, i;
808	struct fio_file *f;
809
810	if (td->io_ops->flags & FIO_CPUIO)
811		return 0;
812
813	fd = open("/dev/urandom", O_RDONLY);
814	if (fd == -1) {
815		td_verror(td, errno);
816		return 1;
817	}
818
819	if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
820		td_verror(td, EIO);
821		close(fd);
822		return 1;
823	}
824
825	close(fd);
826
827	os_random_seed(seeds[0], &td->bsrange_state);
828	os_random_seed(seeds[1], &td->verify_state);
829	os_random_seed(seeds[2], &td->rwmix_state);
830
831	if (td->sequential)
832		return 0;
833
834	if (td->rand_repeatable)
835		seeds[3] = FIO_RANDSEED;
836
837	if (!td->norandommap) {
838		for_each_file(td, f, i) {
839			blocks = (f->file_size + td->rw_min_bs - 1) / td->rw_min_bs;
840			num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
841			f->file_map = malloc(num_maps * sizeof(long));
842			f->num_maps = num_maps;
843			memset(f->file_map, 0, num_maps * sizeof(long));
844		}
845	}
846
847	os_random_seed(seeds[3], &td->random_state);
848	return 0;
849}
850
851static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
852{
853#ifdef FIO_HAVE_CPU_AFFINITY
854	unsigned int i;
855
856	CPU_ZERO(&cpumask);
857
858	for (i = 0; i < sizeof(int) * 8; i++) {
859		if ((1 << i) & cpu)
860			CPU_SET(i, &cpumask);
861	}
862#endif
863}
864
865static int is_empty_or_comment(char *line)
866{
867	unsigned int i;
868
869	for (i = 0; i < strlen(line); i++) {
870		if (line[i] == ';')
871			return 1;
872		if (!isspace(line[i]) && !iscntrl(line[i]))
873			return 0;
874	}
875
876	return 1;
877}
878
879static int str_rw_cb(void *data, const char *mem)
880{
881	struct thread_data *td = data;
882
883	if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
884		td->ddir = DDIR_READ;
885		td->sequential = 1;
886		return 0;
887	} else if (!strncmp(mem, "randread", 8)) {
888		td->ddir = DDIR_READ;
889		td->sequential = 0;
890		return 0;
891	} else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
892		td->ddir = DDIR_WRITE;
893		td->sequential = 1;
894		return 0;
895	} else if (!strncmp(mem, "randwrite", 9)) {
896		td->ddir = DDIR_WRITE;
897		td->sequential = 0;
898		return 0;
899	} else if (!strncmp(mem, "rw", 2)) {
900		td->ddir = DDIR_READ;
901		td->iomix = 1;
902		td->sequential = 1;
903		return 0;
904	} else if (!strncmp(mem, "randrw", 6)) {
905		td->ddir = DDIR_READ;
906		td->iomix = 1;
907		td->sequential = 0;
908		return 0;
909	}
910
911	log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
912	return 1;
913}
914
915static int str_verify_cb(void *data, const char *mem)
916{
917	struct thread_data *td = data;
918
919	if (!strncmp(mem, "0", 1)) {
920		td->verify = VERIFY_NONE;
921		return 0;
922	} else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
923		td->verify = VERIFY_MD5;
924		return 0;
925	} else if (!strncmp(mem, "crc32", 5)) {
926		td->verify = VERIFY_CRC32;
927		return 0;
928	}
929
930	log_err("fio: verify types: md5, crc32\n");
931	return 1;
932}
933
934/*
935 * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
936 */
937static char *get_mmap_file(const char *str)
938{
939	char *p = strstr(str, ":");
940
941	if (!p)
942		return NULL;
943
944	p++;
945	strip_blank_front(&p);
946	strip_blank_end(p);
947	return strdup(p);
948}
949
950static int str_mem_cb(void *data, const char *mem)
951{
952	struct thread_data *td = data;
953
954	if (!strncmp(mem, "malloc", 6)) {
955		td->mem_type = MEM_MALLOC;
956		return 0;
957	} else if (!strncmp(mem, "mmaphuge", 8)) {
958#ifdef FIO_HAVE_HUGETLB
959		/*
960		 * mmaphuge must be appended with the actual file
961		 */
962		td->mmapfile = get_mmap_file(mem);
963		if (!td->mmapfile) {
964			log_err("fio: mmaphuge:/path/to/file\n");
965			return 1;
966		}
967
968		td->mem_type = MEM_MMAPHUGE;
969		return 0;
970#else
971		log_err("fio: mmaphuge not available\n");
972		return 1;
973#endif
974	} else if (!strncmp(mem, "mmap", 4)) {
975		/*
976		 * Check if the user wants file backed memory. It's ok
977		 * if there's no file given, we'll just use anon mamp then.
978		 */
979		td->mmapfile = get_mmap_file(mem);
980		td->mem_type = MEM_MMAP;
981		return 0;
982	} else if (!strncmp(mem, "shmhuge", 7)) {
983#ifdef FIO_HAVE_HUGETLB
984		td->mem_type = MEM_SHMHUGE;
985		return 0;
986#else
987		log_err("fio: shmhuge not available\n");
988		return 1;
989#endif
990	} else if (!strncmp(mem, "shm", 3)) {
991		td->mem_type = MEM_SHM;
992		return 0;
993	}
994
995	log_err("fio: mem type: malloc, shm, shmhuge, mmap, mmaphuge\n");
996	return 1;
997}
998
999static int str_ioengine_cb(void *data, const char *str)
1000{
1001	struct thread_data *td = data;
1002
1003	td->io_ops = load_ioengine(td, str);
1004	if (td->io_ops)
1005		return 0;
1006
1007	log_err("fio: ioengine= libaio, posixaio, sync, mmap, sgio, splice, cpu, null\n");
1008	log_err("fio: or specify path to dynamic ioengine module\n");
1009	return 1;
1010}
1011
1012static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1013{
1014	mlock_size = *val;
1015	return 0;
1016}
1017
1018#ifdef FIO_HAVE_IOPRIO
1019static int str_prioclass_cb(void *data, unsigned int *val)
1020{
1021	struct thread_data *td = data;
1022
1023	td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1024	return 0;
1025}
1026
1027static int str_prio_cb(void *data, unsigned int *val)
1028{
1029	struct thread_data *td = data;
1030
1031	td->ioprio |= *val;
1032	return 0;
1033}
1034#endif
1035
1036static int str_exitall_cb(void)
1037{
1038	exitall_on_terminate = 1;
1039	return 0;
1040}
1041
1042static int str_cpumask_cb(void *data, unsigned int *val)
1043{
1044	struct thread_data *td = data;
1045
1046	fill_cpu_mask(td->cpumask, *val);
1047	return 0;
1048}
1049
1050/*
1051 * This is our [ini] type file parser.
1052 */
1053static int parse_jobs_ini(char *file, int stonewall_flag)
1054{
1055	unsigned int global;
1056	struct thread_data *td;
1057	char *string, *name;
1058	fpos_t off;
1059	FILE *f;
1060	char *p;
1061	int ret = 0, stonewall;
1062
1063	f = fopen(file, "r");
1064	if (!f) {
1065		perror("fopen job file");
1066		return 1;
1067	}
1068
1069	string = malloc(4096);
1070	name = malloc(256);
1071	memset(name, 0, 256);
1072
1073	stonewall = stonewall_flag;
1074	do {
1075		p = fgets(string, 4095, f);
1076		if (!p)
1077			break;
1078		if (is_empty_or_comment(p))
1079			continue;
1080		if (sscanf(p, "[%255s]", name) != 1)
1081			continue;
1082
1083		global = !strncmp(name, "global", 6);
1084
1085		name[strlen(name) - 1] = '\0';
1086
1087		td = get_new_job(global, &def_thread);
1088		if (!td) {
1089			ret = 1;
1090			break;
1091		}
1092
1093		/*
1094		 * Seperate multiple job files by a stonewall
1095		 */
1096		if (!global && stonewall) {
1097			td->stonewall = stonewall;
1098			stonewall = 0;
1099		}
1100
1101		fgetpos(f, &off);
1102		while ((p = fgets(string, 4096, f)) != NULL) {
1103			if (is_empty_or_comment(p))
1104				continue;
1105
1106			strip_blank_front(&p);
1107
1108			if (p[0] == '[')
1109				break;
1110
1111			strip_blank_end(p);
1112
1113			fgetpos(f, &off);
1114
1115			/*
1116			 * Don't break here, continue parsing options so we
1117			 * dump all the bad ones. Makes trial/error fixups
1118			 * easier on the user.
1119			 */
1120			ret |= parse_option(p, options, td);
1121		}
1122
1123		if (!ret) {
1124			fsetpos(f, &off);
1125			ret = add_job(td, name, 0);
1126		} else {
1127			log_err("fio: job %s dropped\n", name);
1128			put_job(td);
1129		}
1130	} while (!ret);
1131
1132	free(string);
1133	free(name);
1134	fclose(f);
1135	return ret;
1136}
1137
1138static int fill_def_thread(void)
1139{
1140	memset(&def_thread, 0, sizeof(def_thread));
1141
1142	if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1143		perror("sched_getaffinity");
1144		return 1;
1145	}
1146
1147	/*
1148	 * fill default options
1149	 */
1150	fill_default_options(&def_thread, options);
1151
1152	def_thread.timeout = def_timeout;
1153	def_thread.write_bw_log = write_bw_log;
1154	def_thread.write_lat_log = write_lat_log;
1155
1156#ifdef FIO_HAVE_DISK_UTIL
1157	def_thread.do_disk_util = 1;
1158#endif
1159
1160	return 0;
1161}
1162
1163static void usage(void)
1164{
1165	printf("%s\n", fio_version_string);
1166	printf("\t--output\tWrite output to file\n");
1167	printf("\t--timeout\tRuntime in seconds\n");
1168	printf("\t--latency-log\tGenerate per-job latency logs\n");
1169	printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1170	printf("\t--minimal\tMinimal (terse) output\n");
1171	printf("\t--version\tPrint version info and exit\n");
1172	printf("\t--help\t\tPrint this page\n");
1173	printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1174}
1175
1176static int parse_cmd_line(int argc, char *argv[])
1177{
1178	struct thread_data *td = NULL;
1179	int c, ini_idx = 0, lidx, ret;
1180
1181	while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1182		switch (c) {
1183		case 't':
1184			def_timeout = atoi(optarg);
1185			break;
1186		case 'l':
1187			write_lat_log = 1;
1188			break;
1189		case 'w':
1190			write_bw_log = 1;
1191			break;
1192		case 'o':
1193			f_out = fopen(optarg, "w+");
1194			if (!f_out) {
1195				perror("fopen output");
1196				exit(1);
1197			}
1198			f_err = f_out;
1199			break;
1200		case 'm':
1201			terse_output = 1;
1202			break;
1203		case 'h':
1204			usage();
1205			exit(0);
1206		case 'c':
1207			ret = show_cmd_help(options, optarg);
1208			exit(ret);
1209		case 'v':
1210			printf("%s\n", fio_version_string);
1211			exit(0);
1212		case FIO_GETOPT_JOB: {
1213			const char *opt = long_options[lidx].name;
1214			char *val = optarg;
1215
1216			if (!strncmp(opt, "name", 4) && td) {
1217				ret = add_job(td, td->name ?: "fio", 0);
1218				if (ret) {
1219					put_job(td);
1220					return 0;
1221				}
1222				td = NULL;
1223			}
1224			if (!td) {
1225				int global = !strncmp(val, "global", 6);
1226
1227				td = get_new_job(global, &def_thread);
1228				if (!td)
1229					return 0;
1230			}
1231
1232			ret = parse_cmd_option(opt, val, options, td);
1233			if (ret) {
1234				log_err("fio: job dropped\n");
1235				put_job(td);
1236				td = NULL;
1237			}
1238			break;
1239		}
1240		default:
1241			break;
1242		}
1243	}
1244
1245	if (td) {
1246		ret = add_job(td, td->name ?: "fio", 0);
1247		if (ret)
1248			put_job(td);
1249	}
1250
1251	while (optind < argc) {
1252		ini_idx++;
1253		ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1254		ini_file[ini_idx - 1] = strdup(argv[optind]);
1255		optind++;
1256	}
1257
1258	return ini_idx;
1259}
1260
1261static void free_shm(void)
1262{
1263	struct shmid_ds sbuf;
1264
1265	if (threads) {
1266		shmdt((void *) threads);
1267		threads = NULL;
1268		shmctl(shm_id, IPC_RMID, &sbuf);
1269	}
1270}
1271
1272/*
1273 * The thread area is shared between the main process and the job
1274 * threads/processes. So setup a shared memory segment that will hold
1275 * all the job info.
1276 */
1277static int setup_thread_area(void)
1278{
1279	/*
1280	 * 1024 is too much on some machines, scale max_jobs if
1281	 * we get a failure that looks like too large a shm segment
1282	 */
1283	do {
1284		size_t size = max_jobs * sizeof(struct thread_data);
1285
1286		shm_id = shmget(0, size, IPC_CREAT | 0600);
1287		if (shm_id != -1)
1288			break;
1289		if (errno != EINVAL) {
1290			perror("shmget");
1291			break;
1292		}
1293
1294		max_jobs >>= 1;
1295	} while (max_jobs);
1296
1297	if (shm_id == -1)
1298		return 1;
1299
1300	threads = shmat(shm_id, NULL, 0);
1301	if (threads == (void *) -1) {
1302		perror("shmat");
1303		return 1;
1304	}
1305
1306	atexit(free_shm);
1307	return 0;
1308}
1309
1310/*
1311 * Copy the fio options into the long options map, so we mirror
1312 * job and cmd line options.
1313 */
1314static void dupe_job_options(void)
1315{
1316	struct fio_option *o;
1317	unsigned int i;
1318
1319	i = 0;
1320	while (long_options[i].name)
1321		i++;
1322
1323	o = &options[0];
1324	while (o->name) {
1325		long_options[i].name = o->name;
1326		long_options[i].val = FIO_GETOPT_JOB;
1327		if (o->type == FIO_OPT_STR_SET)
1328			long_options[i].has_arg = no_argument;
1329		else
1330			long_options[i].has_arg = required_argument;
1331
1332		i++;
1333		o++;
1334		assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1335	}
1336}
1337
1338int parse_options(int argc, char *argv[])
1339{
1340	int job_files, i;
1341
1342	f_out = stdout;
1343	f_err = stderr;
1344
1345	options_init(options);
1346
1347	dupe_job_options();
1348
1349	if (setup_thread_area())
1350		return 1;
1351	if (fill_def_thread())
1352		return 1;
1353
1354	job_files = parse_cmd_line(argc, argv);
1355
1356	for (i = 0; i < job_files; i++) {
1357		if (fill_def_thread())
1358			return 1;
1359		if (parse_jobs_ini(ini_file[i], i))
1360			return 1;
1361		free(ini_file[i]);
1362	}
1363
1364	free(ini_file);
1365
1366	if (!thread_number) {
1367		log_err("No jobs defined(s)\n");
1368		return 1;
1369	}
1370
1371	return 0;
1372}
1373