init.c revision 9728ce37ea4348d4bdb2eefbaafee75ad651758c 
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_ioengine_cb(void *, const char *);
26static int str_mem_cb(void *, const char *);
27static int str_lockmem_cb(void *, unsigned long *);
28#ifdef FIO_HAVE_IOPRIO
29static int str_prio_cb(void *, unsigned int *);
30static int str_prioclass_cb(void *, unsigned int *);
31#endif
32static int str_exitall_cb(void);
33static int str_cpumask_cb(void *, unsigned int *);
34
35#define __stringify_1(x)	#x
36#define __stringify(x)		__stringify_1(x)
37
38/*
39 * Map of job/command line options
40 */
41static struct fio_option options[] = {
42	{
43		.name	= "description",
44		.type	= FIO_OPT_STR_STORE,
45		.off1	= td_var_offset(description),
46		.help	= "Text job description",
47	},
48	{
49		.name	= "name",
50		.type	= FIO_OPT_STR_STORE,
51		.off1	= td_var_offset(name),
52		.help	= "Name of this job",
53	},
54	{
55		.name	= "directory",
56		.type	= FIO_OPT_STR_STORE,
57		.off1	= td_var_offset(directory),
58		.help	= "Directory to store files in",
59	},
60	{
61		.name	= "filename",
62		.type	= FIO_OPT_STR_STORE,
63		.off1	= td_var_offset(filename),
64		.help	= "Force the use of a specific file",
65	},
66	{
67		.name	= "rw",
68		.type	= FIO_OPT_STR,
69		.off1	= td_var_offset(td_ddir),
70		.help	= "IO direction",
71		.def	= "read",
72		.posval = {
73			  { .ival = "read", .oval = TD_DDIR_READ },
74			  { .ival = "write", .oval = TD_DDIR_WRITE },
75			  { .ival = "randread", .oval = TD_DDIR_RANDREAD },
76			  { .ival = "randwrite", .oval = TD_DDIR_RANDWRITE },
77			  { .ival = "rw", .oval = TD_DDIR_RW },
78			  { .ival = "randrw", .oval = TD_DDIR_RANDRW },
79			  },
80	},
81	{
82		.name	= "ioengine",
83		.type	= FIO_OPT_STR,
84		.cb	= str_ioengine_cb,
85		.help	= "IO engine to use",
86		.def	= "sync",
87		.posval	= {
88			  { .ival = "sync", },
89#ifdef FIO_HAVE_LIBAIO
90			  { .ival = "libaio", },
91#endif
92#ifdef FIO_HAVE_POSIXAIO
93			  { .ival = "posixaio", },
94#endif
95			  { .ival = "mmap", },
96#ifdef FIO_HAVE_SPLICE
97			  { .ival = "splice", },
98#endif
99#ifdef FIO_HAVE_SGIO
100			  { .ival = "sg", },
101#endif
102			  { .ival = "null", }, { .ival = "net", },
103#ifdef FIO_HAVE_SYSLET
104			  { .ival = "syslet-rw", },
105#endif
106			  { .ival = "external", },
107			  },
108	},
109	{
110		.name	= "iodepth",
111		.type	= FIO_OPT_INT,
112		.off1	= td_var_offset(iodepth),
113		.help	= "Amount of IO buffers to keep in flight",
114		.def	= "1",
115	},
116	{
117		.name	= "iodepth_batch",
118		.type	= FIO_OPT_INT,
119		.off1	= td_var_offset(iodepth_batch),
120		.help	= "Number of IO to submit in one go",
121	},
122	{
123		.name	= "iodepth_low",
124		.type	= FIO_OPT_INT,
125		.off1	= td_var_offset(iodepth_low),
126		.help	= "Low water mark for queuing depth",
127	},
128	{
129		.name	= "size",
130		.type	= FIO_OPT_STR_VAL,
131		.off1	= td_var_offset(total_file_size),
132		.help	= "Size of device or file",
133	},
134	{
135		.name	= "bs",
136		.type	= FIO_OPT_STR_VAL_INT,
137		.off1	= td_var_offset(bs[DDIR_READ]),
138		.off2	= td_var_offset(bs[DDIR_WRITE]),
139		.help	= "Block size unit",
140		.def	= "4k",
141	},
142	{
143		.name	= "bsrange",
144		.type	= FIO_OPT_RANGE,
145		.off1	= td_var_offset(min_bs[DDIR_READ]),
146		.off2	= td_var_offset(max_bs[DDIR_READ]),
147		.off3	= td_var_offset(min_bs[DDIR_WRITE]),
148		.off4	= td_var_offset(max_bs[DDIR_WRITE]),
149		.help	= "Set block size range (in more detail than bs)",
150	},
151	{
152		.name	= "bs_unaligned",
153		.type	= FIO_OPT_STR_SET,
154		.off1	= td_var_offset(bs_unaligned),
155		.help	= "Don't sector align IO buffer sizes",
156	},
157	{
158		.name	= "offset",
159		.type	= FIO_OPT_STR_VAL,
160		.off1	= td_var_offset(start_offset),
161		.help	= "Start IO from this offset",
162		.def	= "0",
163	},
164	{
165		.name	= "randrepeat",
166		.type	= FIO_OPT_BOOL,
167		.off1	= td_var_offset(rand_repeatable),
168		.help	= "Use repeatable random IO pattern",
169		.def	= "1",
170	},
171	{
172		.name	= "norandommap",
173		.type	= FIO_OPT_STR_SET,
174		.off1	= td_var_offset(norandommap),
175		.help	= "Accept potential duplicate random blocks",
176	},
177	{
178		.name	= "nrfiles",
179		.type	= FIO_OPT_INT,
180		.off1	= td_var_offset(nr_files),
181		.help	= "Split job workload between this number of files",
182		.def	= "1",
183	},
184	{
185		.name	= "file_service_type",
186		.type	= FIO_OPT_STR,
187		.off1	= td_var_offset(file_service_type),
188		.help	= "How to select which file to service next",
189		.def	= "roundrobin",
190		.posval	= {
191			  { .ival = "random", .oval = FIO_FSERVICE_RANDOM },
192			  { .ival = "roundrobin", .oval = FIO_FSERVICE_RR },
193			  },
194	},
195	{
196		.name	= "fsync",
197		.type	= FIO_OPT_INT,
198		.off1	= td_var_offset(fsync_blocks),
199		.help	= "Issue fsync for writes every given number of blocks",
200		.def	= "0",
201	},
202	{
203		.name	= "direct",
204		.type	= FIO_OPT_BOOL,
205		.off1	= td_var_offset(odirect),
206		.help	= "Use O_DIRECT IO (negates buffered)",
207		.def	= "0",
208	},
209	{
210		.name	= "buffered",
211		.type	= FIO_OPT_BOOL,
212		.off1	= td_var_offset(odirect),
213		.neg	= 1,
214		.help	= "Use buffered IO (negates direct)",
215		.def	= "1",
216	},
217	{
218		.name	= "overwrite",
219		.type	= FIO_OPT_BOOL,
220		.off1	= td_var_offset(overwrite),
221		.help	= "When writing, set whether to overwrite current data",
222		.def	= "0",
223	},
224	{
225		.name	= "loops",
226		.type	= FIO_OPT_INT,
227		.off1	= td_var_offset(loops),
228		.help	= "Number of times to run the job",
229		.def	= "1",
230	},
231	{
232		.name	= "numjobs",
233		.type	= FIO_OPT_INT,
234		.off1	= td_var_offset(numjobs),
235		.help	= "Duplicate this job this many times",
236		.def	= "1",
237	},
238	{
239		.name	= "startdelay",
240		.type	= FIO_OPT_INT,
241		.off1	= td_var_offset(start_delay),
242		.help	= "Only start job when this period has passed",
243		.def	= "0",
244	},
245	{
246		.name	= "runtime",
247		.alias	= "timeout",
248		.type	= FIO_OPT_STR_VAL_TIME,
249		.off1	= td_var_offset(timeout),
250		.help	= "Stop workload when this amount of time has passed",
251		.def	= "0",
252	},
253	{
254		.name	= "mem",
255		.type	= FIO_OPT_STR,
256		.cb	= str_mem_cb,
257		.off1	= td_var_offset(mem_type),
258		.help	= "Backing type for IO buffers",
259		.def	= "malloc",
260		.posval	= {
261			  { .ival = "malloc", .oval = MEM_MALLOC },
262			  { .ival = "shm", .oval = MEM_SHM },
263#ifdef FIO_HAVE_HUGETLB
264			  { .ival = "shmhuge", .oval = MEM_SHMHUGE },
265#endif
266			  { .ival = "mmap", .oval = MEM_MMAP },
267#ifdef FIO_HAVE_HUGETLB
268			  { .ival = "mmaphuge", .oval = MEM_MMAPHUGE },
269#endif
270			  },
271	},
272	{
273		.name	= "verify",
274		.type	= FIO_OPT_STR,
275		.off1	= td_var_offset(verify),
276		.help	= "Verify sum function",
277		.def	= "0",
278		.posval = {
279			  { .ival = "0", .oval = VERIFY_NONE },
280			  { .ival = "crc32", .oval = VERIFY_CRC32 },
281			  { .ival = "md5", .oval = VERIFY_MD5 },
282			  },
283	},
284	{
285		.name	= "write_iolog",
286		.type	= FIO_OPT_STR_STORE,
287		.off1	= td_var_offset(write_iolog_file),
288		.help	= "Store IO pattern to file",
289	},
290	{
291		.name	= "read_iolog",
292		.type	= FIO_OPT_STR_STORE,
293		.off1	= td_var_offset(read_iolog_file),
294		.help	= "Playback IO pattern from file",
295	},
296	{
297		.name	= "exec_prerun",
298		.type	= FIO_OPT_STR_STORE,
299		.off1	= td_var_offset(exec_prerun),
300		.help	= "Execute this file prior to running job",
301	},
302	{
303		.name	= "exec_postrun",
304		.type	= FIO_OPT_STR_STORE,
305		.off1	= td_var_offset(exec_postrun),
306		.help	= "Execute this file after running job",
307	},
308#ifdef FIO_HAVE_IOSCHED_SWITCH
309	{
310		.name	= "ioscheduler",
311		.type	= FIO_OPT_STR_STORE,
312		.off1	= td_var_offset(ioscheduler),
313		.help	= "Use this IO scheduler on the backing device",
314	},
315#endif
316	{
317		.name	= "zonesize",
318		.type	= FIO_OPT_STR_VAL,
319		.off1	= td_var_offset(zone_size),
320		.help	= "Give size of an IO zone",
321		.def	= "0",
322	},
323	{
324		.name	= "zoneskip",
325		.type	= FIO_OPT_STR_VAL,
326		.off1	= td_var_offset(zone_skip),
327		.help	= "Space between IO zones",
328		.def	= "0",
329	},
330	{
331		.name	= "lockmem",
332		.type	= FIO_OPT_STR_VAL,
333		.cb	= str_lockmem_cb,
334		.help	= "Lock down this amount of memory",
335		.def	= "0",
336	},
337	{
338		.name	= "rwmixcycle",
339		.type	= FIO_OPT_INT,
340		.off1	= td_var_offset(rwmixcycle),
341		.help	= "Cycle period for mixed read/write workloads (msec)",
342		.def	= "500",
343	},
344	{
345		.name	= "rwmixread",
346		.type	= FIO_OPT_INT,
347		.off1	= td_var_offset(rwmixread),
348		.maxval	= 100,
349		.help	= "Percentage of mixed workload that is reads",
350		.def	= "50",
351	},
352	{
353		.name	= "rwmixwrite",
354		.type	= FIO_OPT_INT,
355		.off1	= td_var_offset(rwmixwrite),
356		.maxval	= 100,
357		.help	= "Percentage of mixed workload that is writes",
358		.def	= "50",
359	},
360	{
361		.name	= "nice",
362		.type	= FIO_OPT_INT,
363		.off1	= td_var_offset(nice),
364		.help	= "Set job CPU nice value",
365		.minval	= -19,
366		.maxval	= 20,
367		.def	= "0",
368	},
369#ifdef FIO_HAVE_IOPRIO
370	{
371		.name	= "prio",
372		.type	= FIO_OPT_INT,
373		.cb	= str_prio_cb,
374		.help	= "Set job IO priority value",
375		.minval	= 0,
376		.maxval	= 7,
377	},
378	{
379		.name	= "prioclass",
380		.type	= FIO_OPT_INT,
381		.cb	= str_prioclass_cb,
382		.help	= "Set job IO priority class",
383		.minval	= 0,
384		.maxval	= 3,
385	},
386#endif
387	{
388		.name	= "thinktime",
389		.type	= FIO_OPT_INT,
390		.off1	= td_var_offset(thinktime),
391		.help	= "Idle time between IO buffers (usec)",
392		.def	= "0",
393	},
394	{
395		.name	= "thinktime_spin",
396		.type	= FIO_OPT_INT,
397		.off1	= td_var_offset(thinktime_spin),
398		.help	= "Start thinktime by spinning this amount (usec)",
399		.def	= "0",
400	},
401	{
402		.name	= "thinktime_blocks",
403		.type	= FIO_OPT_INT,
404		.off1	= td_var_offset(thinktime_blocks),
405		.help	= "IO buffer period between 'thinktime'",
406		.def	= "1",
407	},
408	{
409		.name	= "rate",
410		.type	= FIO_OPT_INT,
411		.off1	= td_var_offset(rate),
412		.help	= "Set bandwidth rate",
413	},
414	{
415		.name	= "ratemin",
416		.type	= FIO_OPT_INT,
417		.off1	= td_var_offset(ratemin),
418		.help	= "The bottom limit accepted",
419	},
420	{
421		.name	= "ratecycle",
422		.type	= FIO_OPT_INT,
423		.off1	= td_var_offset(ratecycle),
424		.help	= "Window average for rate limits (msec)",
425		.def	= "1000",
426	},
427	{
428		.name	= "invalidate",
429		.type	= FIO_OPT_BOOL,
430		.off1	= td_var_offset(invalidate_cache),
431		.help	= "Invalidate buffer/page cache prior to running job",
432		.def	= "1",
433	},
434	{
435		.name	= "sync",
436		.type	= FIO_OPT_BOOL,
437		.off1	= td_var_offset(sync_io),
438		.help	= "Use O_SYNC for buffered writes",
439		.def	= "0",
440	},
441	{
442		.name	= "bwavgtime",
443		.type	= FIO_OPT_INT,
444		.off1	= td_var_offset(bw_avg_time),
445		.help	= "Time window over which to calculate bandwidth (msec)",
446		.def	= "500",
447	},
448	{
449		.name	= "create_serialize",
450		.type	= FIO_OPT_BOOL,
451		.off1	= td_var_offset(create_serialize),
452		.help	= "Serialize creating of job files",
453		.def	= "1",
454	},
455	{
456		.name	= "create_fsync",
457		.type	= FIO_OPT_BOOL,
458		.off1	= td_var_offset(create_fsync),
459		.help	= "Fsync file after creation",
460		.def	= "1",
461	},
462	{
463		.name	= "cpuload",
464		.type	= FIO_OPT_INT,
465		.off1	= td_var_offset(cpuload),
466		.help	= "Use this percentage of CPU",
467	},
468	{
469		.name	= "cpuchunks",
470		.type	= FIO_OPT_INT,
471		.off1	= td_var_offset(cpucycle),
472		.help	= "Length of the CPU burn cycles",
473	},
474#ifdef FIO_HAVE_CPU_AFFINITY
475	{
476		.name	= "cpumask",
477		.type	= FIO_OPT_INT,
478		.cb	= str_cpumask_cb,
479		.help	= "CPU affinity mask",
480	},
481#endif
482	{
483		.name	= "end_fsync",
484		.type	= FIO_OPT_BOOL,
485		.off1	= td_var_offset(end_fsync),
486		.help	= "Include fsync at the end of job",
487		.def	= "0",
488	},
489	{
490		.name	= "unlink",
491		.type	= FIO_OPT_BOOL,
492		.off1	= td_var_offset(unlink),
493		.help	= "Unlink created files after job has completed",
494		.def	= "0",
495	},
496	{
497		.name	= "exitall",
498		.type	= FIO_OPT_STR_SET,
499		.cb	= str_exitall_cb,
500		.help	= "Terminate all jobs when one exits",
501	},
502	{
503		.name	= "stonewall",
504		.type	= FIO_OPT_STR_SET,
505		.off1	= td_var_offset(stonewall),
506		.help	= "Insert a hard barrier between this job and previous",
507	},
508	{
509		.name	= "thread",
510		.type	= FIO_OPT_STR_SET,
511		.off1	= td_var_offset(use_thread),
512		.help	= "Use threads instead of forks",
513	},
514	{
515		.name	= "write_bw_log",
516		.type	= FIO_OPT_STR_SET,
517		.off1	= td_var_offset(write_bw_log),
518		.help	= "Write log of bandwidth during run",
519	},
520	{
521		.name	= "write_lat_log",
522		.type	= FIO_OPT_STR_SET,
523		.off1	= td_var_offset(write_lat_log),
524		.help	= "Write log of latency during run",
525	},
526	{
527		.name	= "hugepage-size",
528		.type	= FIO_OPT_STR_VAL,
529		.off1	= td_var_offset(hugepage_size),
530		.help	= "When using hugepages, specify size of each page",
531		.def	= __stringify(FIO_HUGE_PAGE),
532	},
533	{
534		.name = NULL,
535	},
536};
537
538#define FIO_JOB_OPTS	(sizeof(options) / sizeof(struct fio_option))
539#define FIO_CMD_OPTS	(16)
540#define FIO_GETOPT_JOB	(0x89988998)
541
542/*
543 * Command line options. These will contain the above, plus a few
544 * extra that only pertain to fio itself and not jobs.
545 */
546static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
547	{
548		.name		= "output",
549		.has_arg	= required_argument,
550		.val		= 'o',
551	},
552	{
553		.name		= "timeout",
554		.has_arg	= required_argument,
555		.val		= 't',
556	},
557	{
558		.name		= "latency-log",
559		.has_arg	= required_argument,
560		.val		= 'l',
561	},
562	{
563		.name		= "bandwidth-log",
564		.has_arg	= required_argument,
565		.val		= 'b',
566	},
567	{
568		.name		= "minimal",
569		.has_arg	= optional_argument,
570		.val		= 'm',
571	},
572	{
573		.name		= "version",
574		.has_arg	= no_argument,
575		.val		= 'v',
576	},
577	{
578		.name		= "help",
579		.has_arg	= no_argument,
580		.val		= 'h',
581	},
582	{
583		.name		= "cmdhelp",
584		.has_arg	= required_argument,
585		.val		= 'c',
586	},
587	{
588		.name		= NULL,
589	},
590};
591
592static int def_timeout = 0;
593
594static char fio_version_string[] = "fio 1.12";
595
596static char **ini_file;
597static int max_jobs = MAX_JOBS;
598
599struct thread_data def_thread;
600struct thread_data *threads = NULL;
601
602int exitall_on_terminate = 0;
603int terse_output = 0;
604unsigned long long mlock_size = 0;
605FILE *f_out = NULL;
606FILE *f_err = NULL;
607
608static int write_lat_log = 0;
609int write_bw_log = 0;
610
611FILE *get_f_out()
612{
613	return f_out;
614}
615
616FILE *get_f_err()
617{
618	return f_err;
619}
620
621/*
622 * Return a free job structure.
623 */
624static struct thread_data *get_new_job(int global, struct thread_data *parent)
625{
626	struct thread_data *td;
627
628	if (global)
629		return &def_thread;
630	if (thread_number >= max_jobs)
631		return NULL;
632
633	td = &threads[thread_number++];
634	*td = *parent;
635
636	td->thread_number = thread_number;
637	return td;
638}
639
640static void put_job(struct thread_data *td)
641{
642	if (td == &def_thread)
643		return;
644
645	if (td->error)
646		fprintf(f_out, "fio: %s\n", td->verror);
647
648	memset(&threads[td->thread_number - 1], 0, sizeof(*td));
649	thread_number--;
650}
651
652/*
653 * Lazy way of fixing up options that depend on each other. We could also
654 * define option callback handlers, but this is easier.
655 */
656static void fixup_options(struct thread_data *td)
657{
658	if (!td->rwmixread && td->rwmixwrite)
659		td->rwmixread = 100 - td->rwmixwrite;
660
661	if (td->write_iolog_file && td->read_iolog_file) {
662		log_err("fio: read iolog overrides write_iolog\n");
663		free(td->write_iolog_file);
664		td->write_iolog_file = NULL;
665	}
666
667	if (td->io_ops->flags & FIO_SYNCIO)
668		td->iodepth = 1;
669	else {
670		if (!td->iodepth)
671			td->iodepth = td->nr_files;
672	}
673
674	/*
675	 * only really works for sequential io for now, and with 1 file
676	 */
677	if (td->zone_size && td_random(td) && td->nr_files == 1)
678		td->zone_size = 0;
679
680	/*
681	 * Reads can do overwrites, we always need to pre-create the file
682	 */
683	if (td_read(td) || td_rw(td))
684		td->overwrite = 1;
685
686	if (!td->min_bs[DDIR_READ])
687		td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
688	if (!td->max_bs[DDIR_READ])
689		td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
690	if (!td->min_bs[DDIR_WRITE])
691		td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
692	if (!td->max_bs[DDIR_WRITE])
693		td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
694
695	td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
696
697	if (td_read(td) && !td_rw(td))
698		td->verify = 0;
699
700	if (td->norandommap && td->verify != VERIFY_NONE) {
701		log_err("fio: norandommap given, verify disabled\n");
702		td->verify = VERIFY_NONE;
703	}
704	if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
705		log_err("fio: bs_unaligned may not work with raw io\n");
706
707	/*
708	 * O_DIRECT and char doesn't mix, clear that flag if necessary.
709	 */
710	if (td->filetype == FIO_TYPE_CHAR && td->odirect)
711		td->odirect = 0;
712
713	/*
714	 * thinktime_spin must be less than thinktime
715	 */
716	if (td->thinktime_spin > td->thinktime)
717		td->thinktime_spin = td->thinktime;
718
719	/*
720	 * The low water mark cannot be bigger than the iodepth
721	 */
722	if (td->iodepth_low > td->iodepth || !td->iodepth_low) {
723		/*
724		 * syslet work around - if the workload is sequential,
725		 * we want to let the queue drain all the way down to
726		 * avoid seeking between async threads
727		 */
728		if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
729			td->iodepth_low = 1;
730		else
731			td->iodepth_low = td->iodepth;
732	}
733
734	/*
735	 * If batch number isn't set, default to the same as iodepth
736	 */
737	if (td->iodepth_batch > td->iodepth || !td->iodepth_batch)
738		td->iodepth_batch = td->iodepth;
739}
740
741/*
742 * This function leaks the buffer
743 */
744static char *to_kmg(unsigned int val)
745{
746	char *buf = malloc(32);
747	char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
748	char *p = post;
749
750	do {
751		if (val & 1023)
752			break;
753
754		val >>= 10;
755		p++;
756	} while (*p);
757
758	snprintf(buf, 31, "%u%c", val, *p);
759	return buf;
760}
761
762/*
763 * Adds a job to the list of things todo. Sanitizes the various options
764 * to make sure we don't have conflicts, and initializes various
765 * members of td.
766 */
767static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
768{
769	const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
770				   "randread", "randwrite", "randrw" };
771	struct stat sb;
772	int numjobs, i;
773	struct fio_file *f;
774
775	/*
776	 * the def_thread is just for options, it's not a real job
777	 */
778	if (td == &def_thread)
779		return 0;
780
781	assert(td->io_ops);
782
783	if (td->odirect)
784		td->io_ops->flags |= FIO_RAWIO;
785
786	td->filetype = FIO_TYPE_FILE;
787	if (td->filename && !lstat(td->filename, &sb)) {
788		if (S_ISBLK(sb.st_mode))
789			td->filetype = FIO_TYPE_BD;
790		else if (S_ISCHR(sb.st_mode))
791			td->filetype = FIO_TYPE_CHAR;
792	}
793
794	fixup_options(td);
795
796	if (td->filename)
797		td->nr_uniq_files = 1;
798	else
799		td->nr_uniq_files = td->nr_files;
800
801	if (td->filetype == FIO_TYPE_FILE || td->filename) {
802		char tmp[PATH_MAX];
803		int len = 0;
804
805		if (td->directory && td->directory[0] != '\0') {
806			if (lstat(td->directory, &sb) < 0) {
807				log_err("fio: %s is not a directory\n", td->directory);
808				td_verror(td, errno, "lstat");
809				return 1;
810			}
811			if (!S_ISDIR(sb.st_mode)) {
812				log_err("fio: %s is not a directory\n", td->directory);
813				return 1;
814			}
815			len = sprintf(tmp, "%s/", td->directory);
816		}
817
818		td->files = malloc(sizeof(struct fio_file) * td->nr_files);
819
820		for_each_file(td, f, i) {
821			memset(f, 0, sizeof(*f));
822			f->fd = -1;
823
824			if (td->filename)
825				sprintf(tmp + len, "%s", td->filename);
826			else
827				sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
828			f->file_name = strdup(tmp);
829		}
830	} else {
831		td->nr_files = 1;
832		td->files = malloc(sizeof(struct fio_file));
833		f = &td->files[0];
834
835		memset(f, 0, sizeof(*f));
836		f->fd = -1;
837		f->file_name = strdup(jobname);
838	}
839
840	for_each_file(td, f, i) {
841		f->file_size = td->total_file_size / td->nr_files;
842		f->file_offset = td->start_offset;
843	}
844
845	fio_sem_init(&td->mutex, 0);
846
847	td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
848	td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
849	td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
850
851	if (td->stonewall && td->thread_number > 1)
852		groupid++;
853
854	td->groupid = groupid;
855
856	if (setup_rate(td))
857		goto err;
858
859	if (td->write_lat_log) {
860		setup_log(&td->ts.slat_log);
861		setup_log(&td->ts.clat_log);
862	}
863	if (td->write_bw_log)
864		setup_log(&td->ts.bw_log);
865
866	if (!td->name)
867		td->name = strdup(jobname);
868
869	if (!terse_output) {
870		if (!job_add_num) {
871			if (td->io_ops->flags & FIO_CPUIO)
872				fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
873			else {
874				char *c1, *c2, *c3, *c4;
875
876				c1 = to_kmg(td->min_bs[DDIR_READ]);
877				c2 = to_kmg(td->max_bs[DDIR_READ]);
878				c3 = to_kmg(td->min_bs[DDIR_WRITE]);
879				c4 = to_kmg(td->max_bs[DDIR_WRITE]);
880
881				fprintf(f_out, "%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[td->td_ddir], c1, c2, c3, c4, td->io_ops->name, td->iodepth);
882
883				free(c1);
884				free(c2);
885				free(c3);
886				free(c4);
887			}
888		} else if (job_add_num == 1)
889			fprintf(f_out, "...\n");
890	}
891
892	/*
893	 * recurse add identical jobs, clear numjobs and stonewall options
894	 * as they don't apply to sub-jobs
895	 */
896	numjobs = td->numjobs;
897	while (--numjobs) {
898		struct thread_data *td_new = get_new_job(0, td);
899
900		if (!td_new)
901			goto err;
902
903		td_new->numjobs = 1;
904		td_new->stonewall = 0;
905		job_add_num = numjobs - 1;
906
907		if (add_job(td_new, jobname, job_add_num))
908			goto err;
909	}
910	return 0;
911err:
912	put_job(td);
913	return -1;
914}
915
916/*
917 * Initialize the various random states we need (random io, block size ranges,
918 * read/write mix, etc).
919 */
920int init_random_state(struct thread_data *td)
921{
922	unsigned long seeds[5];
923	int fd, num_maps, blocks, i;
924	struct fio_file *f;
925
926	if (td->io_ops->flags & FIO_CPUIO)
927		return 0;
928
929	fd = open("/dev/urandom", O_RDONLY);
930	if (fd == -1) {
931		td_verror(td, errno, "open");
932		return 1;
933	}
934
935	if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
936		td_verror(td, EIO, "read");
937		close(fd);
938		return 1;
939	}
940
941	close(fd);
942
943	os_random_seed(seeds[0], &td->bsrange_state);
944	os_random_seed(seeds[1], &td->verify_state);
945	os_random_seed(seeds[2], &td->rwmix_state);
946
947	if (td->file_service_type == FIO_FSERVICE_RANDOM)
948		os_random_seed(seeds[3], &td->next_file_state);
949
950	if (!td_random(td))
951		return 0;
952
953	if (td->rand_repeatable)
954		seeds[4] = FIO_RANDSEED * td->thread_number;
955
956	if (!td->norandommap) {
957		for_each_file(td, f, i) {
958			blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
959			num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
960			f->file_map = malloc(num_maps * sizeof(long));
961			f->num_maps = num_maps;
962			memset(f->file_map, 0, num_maps * sizeof(long));
963		}
964	}
965
966	os_random_seed(seeds[4], &td->random_state);
967	return 0;
968}
969
970static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
971{
972#ifdef FIO_HAVE_CPU_AFFINITY
973	unsigned int i;
974
975	CPU_ZERO(&cpumask);
976
977	for (i = 0; i < sizeof(int) * 8; i++) {
978		if ((1 << i) & cpu)
979			CPU_SET(i, &cpumask);
980	}
981#endif
982}
983
984static int is_empty_or_comment(char *line)
985{
986	unsigned int i;
987
988	for (i = 0; i < strlen(line); i++) {
989		if (line[i] == ';')
990			return 1;
991		if (line[i] == '#')
992			return 1;
993		if (!isspace(line[i]) && !iscntrl(line[i]))
994			return 0;
995	}
996
997	return 1;
998}
999
1000/*
1001 * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
1002 */
1003static char *get_mmap_file(const char *str)
1004{
1005	char *p = strstr(str, ":");
1006
1007	if (!p)
1008		return NULL;
1009
1010	p++;
1011	strip_blank_front(&p);
1012	strip_blank_end(p);
1013	return strdup(p);
1014}
1015
1016static int str_mem_cb(void *data, const char *mem)
1017{
1018	struct thread_data *td = data;
1019
1020	if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
1021		td->mmapfile = get_mmap_file(mem);
1022		if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
1023			log_err("fio: mmaphuge:/path/to/file\n");
1024			return 1;
1025		}
1026	}
1027
1028	return 0;
1029}
1030
1031/* External engines are specified by "external:name.o") */
1032static const char *get_engine_name(const char *str)
1033{
1034	char *p = strstr(str, ":");
1035
1036	if (!p)
1037		return str;
1038
1039	p++;
1040	strip_blank_front(&p);
1041	strip_blank_end(p);
1042	return p;
1043}
1044
1045static int str_ioengine_cb(void *data, const char *str)
1046{
1047	struct thread_data *td = data;
1048	const char *name = get_engine_name(str);
1049
1050	td->io_ops = load_ioengine(td, name);
1051	if (td->io_ops)
1052		return 0;
1053
1054	return 1;
1055}
1056
1057static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1058{
1059	mlock_size = *val;
1060	return 0;
1061}
1062
1063#ifdef FIO_HAVE_IOPRIO
1064static int str_prioclass_cb(void *data, unsigned int *val)
1065{
1066	struct thread_data *td = data;
1067
1068	td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1069	return 0;
1070}
1071
1072static int str_prio_cb(void *data, unsigned int *val)
1073{
1074	struct thread_data *td = data;
1075
1076	td->ioprio |= *val;
1077	return 0;
1078}
1079#endif
1080
1081static int str_exitall_cb(void)
1082{
1083	exitall_on_terminate = 1;
1084	return 0;
1085}
1086
1087static int str_cpumask_cb(void *data, unsigned int *val)
1088{
1089	struct thread_data *td = data;
1090
1091	fill_cpu_mask(td->cpumask, *val);
1092	return 0;
1093}
1094
1095/*
1096 * This is our [ini] type file parser.
1097 */
1098static int parse_jobs_ini(char *file, int stonewall_flag)
1099{
1100	unsigned int global;
1101	struct thread_data *td;
1102	char *string, *name;
1103	fpos_t off;
1104	FILE *f;
1105	char *p;
1106	int ret = 0, stonewall;
1107
1108	f = fopen(file, "r");
1109	if (!f) {
1110		perror("fopen job file");
1111		return 1;
1112	}
1113
1114	string = malloc(4096);
1115	name = malloc(256);
1116	memset(name, 0, 256);
1117
1118	stonewall = stonewall_flag;
1119	do {
1120		p = fgets(string, 4095, f);
1121		if (!p)
1122			break;
1123		if (is_empty_or_comment(p))
1124			continue;
1125		if (sscanf(p, "[%255s]", name) != 1)
1126			continue;
1127
1128		global = !strncmp(name, "global", 6);
1129
1130		name[strlen(name) - 1] = '\0';
1131
1132		td = get_new_job(global, &def_thread);
1133		if (!td) {
1134			ret = 1;
1135			break;
1136		}
1137
1138		/*
1139		 * Seperate multiple job files by a stonewall
1140		 */
1141		if (!global && stonewall) {
1142			td->stonewall = stonewall;
1143			stonewall = 0;
1144		}
1145
1146		fgetpos(f, &off);
1147		while ((p = fgets(string, 4096, f)) != NULL) {
1148			if (is_empty_or_comment(p))
1149				continue;
1150
1151			strip_blank_front(&p);
1152
1153			if (p[0] == '[')
1154				break;
1155
1156			strip_blank_end(p);
1157
1158			fgetpos(f, &off);
1159
1160			/*
1161			 * Don't break here, continue parsing options so we
1162			 * dump all the bad ones. Makes trial/error fixups
1163			 * easier on the user.
1164			 */
1165			ret |= parse_option(p, options, td);
1166		}
1167
1168		if (!ret) {
1169			fsetpos(f, &off);
1170			ret = add_job(td, name, 0);
1171		} else {
1172			log_err("fio: job %s dropped\n", name);
1173			put_job(td);
1174		}
1175	} while (!ret);
1176
1177	free(string);
1178	free(name);
1179	fclose(f);
1180	return ret;
1181}
1182
1183static int fill_def_thread(void)
1184{
1185	memset(&def_thread, 0, sizeof(def_thread));
1186
1187	if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1188		perror("sched_getaffinity");
1189		return 1;
1190	}
1191
1192	/*
1193	 * fill default options
1194	 */
1195	fill_default_options(&def_thread, options);
1196
1197	def_thread.timeout = def_timeout;
1198	def_thread.write_bw_log = write_bw_log;
1199	def_thread.write_lat_log = write_lat_log;
1200
1201#ifdef FIO_HAVE_DISK_UTIL
1202	def_thread.do_disk_util = 1;
1203#endif
1204
1205	return 0;
1206}
1207
1208static void usage(void)
1209{
1210	printf("%s\n", fio_version_string);
1211	printf("\t--output\tWrite output to file\n");
1212	printf("\t--timeout\tRuntime in seconds\n");
1213	printf("\t--latency-log\tGenerate per-job latency logs\n");
1214	printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1215	printf("\t--minimal\tMinimal (terse) output\n");
1216	printf("\t--version\tPrint version info and exit\n");
1217	printf("\t--help\t\tPrint this page\n");
1218	printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1219}
1220
1221static int parse_cmd_line(int argc, char *argv[])
1222{
1223	struct thread_data *td = NULL;
1224	int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1225
1226	while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1227		switch (c) {
1228		case 't':
1229			def_timeout = atoi(optarg);
1230			break;
1231		case 'l':
1232			write_lat_log = 1;
1233			break;
1234		case 'w':
1235			write_bw_log = 1;
1236			break;
1237		case 'o':
1238			f_out = fopen(optarg, "w+");
1239			if (!f_out) {
1240				perror("fopen output");
1241				exit(1);
1242			}
1243			f_err = f_out;
1244			break;
1245		case 'm':
1246			terse_output = 1;
1247			break;
1248		case 'h':
1249			usage();
1250			exit(0);
1251		case 'c':
1252			ret = show_cmd_help(options, optarg);
1253			exit(ret);
1254		case 'v':
1255			printf("%s\n", fio_version_string);
1256			exit(0);
1257		case FIO_GETOPT_JOB: {
1258			const char *opt = long_options[lidx].name;
1259			char *val = optarg;
1260
1261			if (!strncmp(opt, "name", 4) && td) {
1262				ret = add_job(td, td->name ?: "fio", 0);
1263				if (ret) {
1264					put_job(td);
1265					return 0;
1266				}
1267				td = NULL;
1268			}
1269			if (!td) {
1270				int global = !strncmp(val, "global", 6);
1271
1272				td = get_new_job(global, &def_thread);
1273				if (!td)
1274					return 0;
1275			}
1276
1277			ret = parse_cmd_option(opt, val, options, td);
1278			if (ret)
1279				dont_add_job = 1;
1280			break;
1281		}
1282		default:
1283			break;
1284		}
1285	}
1286
1287	if (td) {
1288		if (dont_add_job)
1289			put_job(td);
1290		else {
1291			ret = add_job(td, td->name ?: "fio", 0);
1292			if (ret)
1293				put_job(td);
1294		}
1295	}
1296
1297	while (optind < argc) {
1298		ini_idx++;
1299		ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1300		ini_file[ini_idx - 1] = strdup(argv[optind]);
1301		optind++;
1302	}
1303
1304	return ini_idx;
1305}
1306
1307static void free_shm(void)
1308{
1309	struct shmid_ds sbuf;
1310
1311	if (threads) {
1312		shmdt((void *) threads);
1313		threads = NULL;
1314		shmctl(shm_id, IPC_RMID, &sbuf);
1315	}
1316}
1317
1318/*
1319 * The thread area is shared between the main process and the job
1320 * threads/processes. So setup a shared memory segment that will hold
1321 * all the job info.
1322 */
1323static int setup_thread_area(void)
1324{
1325	/*
1326	 * 1024 is too much on some machines, scale max_jobs if
1327	 * we get a failure that looks like too large a shm segment
1328	 */
1329	do {
1330		size_t size = max_jobs * sizeof(struct thread_data);
1331
1332		shm_id = shmget(0, size, IPC_CREAT | 0600);
1333		if (shm_id != -1)
1334			break;
1335		if (errno != EINVAL) {
1336			perror("shmget");
1337			break;
1338		}
1339
1340		max_jobs >>= 1;
1341	} while (max_jobs);
1342
1343	if (shm_id == -1)
1344		return 1;
1345
1346	threads = shmat(shm_id, NULL, 0);
1347	if (threads == (void *) -1) {
1348		perror("shmat");
1349		return 1;
1350	}
1351
1352	atexit(free_shm);
1353	return 0;
1354}
1355
1356/*
1357 * Copy the fio options into the long options map, so we mirror
1358 * job and cmd line options.
1359 */
1360static void dupe_job_options(void)
1361{
1362	struct fio_option *o;
1363	unsigned int i;
1364
1365	i = 0;
1366	while (long_options[i].name)
1367		i++;
1368
1369	o = &options[0];
1370	while (o->name) {
1371		long_options[i].name = o->name;
1372		long_options[i].val = FIO_GETOPT_JOB;
1373		if (o->type == FIO_OPT_STR_SET)
1374			long_options[i].has_arg = no_argument;
1375		else
1376			long_options[i].has_arg = required_argument;
1377
1378		i++;
1379		o++;
1380		assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1381	}
1382}
1383
1384int parse_options(int argc, char *argv[])
1385{
1386	int job_files, i;
1387
1388	f_out = stdout;
1389	f_err = stderr;
1390
1391	options_init(options);
1392
1393	dupe_job_options();
1394
1395	if (setup_thread_area())
1396		return 1;
1397	if (fill_def_thread())
1398		return 1;
1399
1400	job_files = parse_cmd_line(argc, argv);
1401
1402	for (i = 0; i < job_files; i++) {
1403		if (fill_def_thread())
1404			return 1;
1405		if (parse_jobs_ini(ini_file[i], i))
1406			return 1;
1407		free(ini_file[i]);
1408	}
1409
1410	free(ini_file);
1411
1412	if (!thread_number) {
1413		log_err("No jobs defined(s)\n");
1414		return 1;
1415	}
1416
1417	return 0;
1418}
1419