init.c revision 413dd459a7710ba421061e840dd9ac3161c70f20 
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 *);
36static int str_file_service_cb(void *, const char *);
37
38#define __stringify_1(x)	#x
39#define __stringify(x)		__stringify_1(x)
40
41/*
42 * Map of job/command line options
43 */
44static struct fio_option options[] = {
45	{
46		.name	= "description",
47		.type	= FIO_OPT_STR_STORE,
48		.off1	= td_var_offset(description),
49		.help	= "Text job description",
50	},
51	{
52		.name	= "name",
53		.type	= FIO_OPT_STR_STORE,
54		.off1	= td_var_offset(name),
55		.help	= "Name of this job",
56	},
57	{
58		.name	= "directory",
59		.type	= FIO_OPT_STR_STORE,
60		.off1	= td_var_offset(directory),
61		.help	= "Directory to store files in",
62	},
63	{
64		.name	= "filename",
65		.type	= FIO_OPT_STR_STORE,
66		.off1	= td_var_offset(filename),
67		.help	= "Force the use of a specific file",
68	},
69	{
70		.name	= "rw",
71		.type	= FIO_OPT_STR,
72		.cb	= str_rw_cb,
73		.help	= "IO direction",
74		.def	= "read",
75		.posval	= { "read", "write", "randwrite", "randread", "rw",
76				"randrw", },
77	},
78	{
79		.name	= "ioengine",
80		.type	= FIO_OPT_STR,
81		.cb	= str_ioengine_cb,
82		.help	= "IO engine to use",
83		.def	= "sync",
84		.posval	= { "sync", "libaio", "posixaio", "mmap", "splice",
85				"sg", "null", "net", "syslet-rw" },
86	},
87	{
88		.name	= "iodepth",
89		.type	= FIO_OPT_INT,
90		.off1	= td_var_offset(iodepth),
91		.help	= "Amount of IO buffers to keep in flight",
92		.def	= "1",
93	},
94	{
95		.name	= "iodepth_low",
96		.type	= FIO_OPT_INT,
97		.off1	= td_var_offset(iodepth_low),
98		.help	= "Low water mark for queuing depth",
99	},
100	{
101		.name	= "size",
102		.type	= FIO_OPT_STR_VAL,
103		.off1	= td_var_offset(total_file_size),
104		.help	= "Size of device or file",
105	},
106	{
107		.name	= "bs",
108		.type	= FIO_OPT_STR_VAL_INT,
109		.off1	= td_var_offset(bs[DDIR_READ]),
110		.off2	= td_var_offset(bs[DDIR_WRITE]),
111		.help	= "Block size unit",
112		.def	= "4k",
113	},
114	{
115		.name	= "bsrange",
116		.type	= FIO_OPT_RANGE,
117		.off1	= td_var_offset(min_bs[DDIR_READ]),
118		.off2	= td_var_offset(max_bs[DDIR_READ]),
119		.off3	= td_var_offset(min_bs[DDIR_WRITE]),
120		.off4	= td_var_offset(max_bs[DDIR_WRITE]),
121		.help	= "Set block size range (in more detail than bs)",
122	},
123	{
124		.name	= "bs_unaligned",
125		.type	= FIO_OPT_STR_SET,
126		.off1	= td_var_offset(bs_unaligned),
127		.help	= "Don't sector align IO buffer sizes",
128	},
129	{
130		.name	= "offset",
131		.type	= FIO_OPT_STR_VAL,
132		.off1	= td_var_offset(start_offset),
133		.help	= "Start IO from this offset",
134		.def	= "0",
135	},
136	{
137		.name	= "randrepeat",
138		.type	= FIO_OPT_BOOL,
139		.off1	= td_var_offset(rand_repeatable),
140		.help	= "Use repeatable random IO pattern",
141		.def	= "1",
142	},
143	{
144		.name	= "norandommap",
145		.type	= FIO_OPT_STR_SET,
146		.off1	= td_var_offset(norandommap),
147		.help	= "Accept potential duplicate random blocks",
148	},
149	{
150		.name	= "nrfiles",
151		.type	= FIO_OPT_INT,
152		.off1	= td_var_offset(nr_files),
153		.help	= "Split job workload between this number of files",
154		.def	= "1",
155	},
156	{
157		.name	= "file_service_type",
158		.type	= FIO_OPT_STR,
159		.cb	= str_file_service_cb,
160		.help	= "How to select which file to service next",
161		.def	= "roundrobin",
162		.posval	= { "random", "roundrobin" },
163	},
164	{
165		.name	= "fsync",
166		.type	= FIO_OPT_INT,
167		.off1	= td_var_offset(fsync_blocks),
168		.help	= "Issue fsync for writes every given number of blocks",
169		.def	= "0",
170	},
171	{
172		.name	= "direct",
173		.type	= FIO_OPT_BOOL,
174		.off1	= td_var_offset(odirect),
175		.help	= "Use O_DIRECT IO (negates buffered)",
176		.def	= "0",
177	},
178	{
179		.name	= "buffered",
180		.type	= FIO_OPT_BOOL,
181		.off1	= td_var_offset(odirect),
182		.neg	= 1,
183		.help	= "Use buffered IO (negates direct)",
184		.def	= "1",
185	},
186	{
187		.name	= "overwrite",
188		.type	= FIO_OPT_BOOL,
189		.off1	= td_var_offset(overwrite),
190		.help	= "When writing, set whether to overwrite current data",
191		.def	= "0",
192	},
193	{
194		.name	= "loops",
195		.type	= FIO_OPT_INT,
196		.off1	= td_var_offset(loops),
197		.help	= "Number of times to run the job",
198		.def	= "1",
199	},
200	{
201		.name	= "numjobs",
202		.type	= FIO_OPT_INT,
203		.off1	= td_var_offset(numjobs),
204		.help	= "Duplicate this job this many times",
205		.def	= "1",
206	},
207	{
208		.name	= "startdelay",
209		.type	= FIO_OPT_INT,
210		.off1	= td_var_offset(start_delay),
211		.help	= "Only start job when this period has passed",
212		.def	= "0",
213	},
214	{
215		.name	= "runtime",
216		.alias	= "timeout",
217		.type	= FIO_OPT_STR_VAL_TIME,
218		.off1	= td_var_offset(timeout),
219		.help	= "Stop workload when this amount of time has passed",
220		.def	= "0",
221	},
222	{
223		.name	= "mem",
224		.type	= FIO_OPT_STR,
225		.cb	= str_mem_cb,
226		.help	= "Backing type for IO buffers",
227		.def	= "malloc",
228		.posval	=  { "malloc", "shm", "shmhuge", "mmap", "mmaphuge", },
229	},
230	{
231		.name	= "verify",
232		.type	= FIO_OPT_STR,
233		.cb	= str_verify_cb,
234		.help	= "Verify sum function",
235		.def	= "0",
236		.posval	= { "crc32", "md5", },
237	},
238	{
239		.name	= "write_iolog",
240		.type	= FIO_OPT_STR_STORE,
241		.off1	= td_var_offset(write_iolog_file),
242		.help	= "Store IO pattern to file",
243	},
244	{
245		.name	= "read_iolog",
246		.type	= FIO_OPT_STR_STORE,
247		.off1	= td_var_offset(read_iolog_file),
248		.help	= "Playback IO pattern from file",
249	},
250	{
251		.name	= "exec_prerun",
252		.type	= FIO_OPT_STR_STORE,
253		.off1	= td_var_offset(exec_prerun),
254		.help	= "Execute this file prior to running job",
255	},
256	{
257		.name	= "exec_postrun",
258		.type	= FIO_OPT_STR_STORE,
259		.off1	= td_var_offset(exec_postrun),
260		.help	= "Execute this file after running job",
261	},
262#ifdef FIO_HAVE_IOSCHED_SWITCH
263	{
264		.name	= "ioscheduler",
265		.type	= FIO_OPT_STR_STORE,
266		.off1	= td_var_offset(ioscheduler),
267		.help	= "Use this IO scheduler on the backing device",
268	},
269#endif
270	{
271		.name	= "zonesize",
272		.type	= FIO_OPT_STR_VAL,
273		.off1	= td_var_offset(zone_size),
274		.help	= "Give size of an IO zone",
275		.def	= "0",
276	},
277	{
278		.name	= "zoneskip",
279		.type	= FIO_OPT_STR_VAL,
280		.off1	= td_var_offset(zone_skip),
281		.help	= "Space between IO zones",
282		.def	= "0",
283	},
284	{
285		.name	= "lockmem",
286		.type	= FIO_OPT_STR_VAL,
287		.cb	= str_lockmem_cb,
288		.help	= "Lock down this amount of memory",
289		.def	= "0",
290	},
291	{
292		.name	= "rwmixcycle",
293		.type	= FIO_OPT_INT,
294		.off1	= td_var_offset(rwmixcycle),
295		.help	= "Cycle period for mixed read/write workloads (msec)",
296		.def	= "500",
297	},
298	{
299		.name	= "rwmixread",
300		.type	= FIO_OPT_INT,
301		.off1	= td_var_offset(rwmixread),
302		.maxval	= 100,
303		.help	= "Percentage of mixed workload that is reads",
304		.def	= "50",
305	},
306	{
307		.name	= "rwmixwrite",
308		.type	= FIO_OPT_INT,
309		.off1	= td_var_offset(rwmixwrite),
310		.maxval	= 100,
311		.help	= "Percentage of mixed workload that is writes",
312		.def	= "50",
313	},
314	{
315		.name	= "nice",
316		.type	= FIO_OPT_INT,
317		.off1	= td_var_offset(nice),
318		.help	= "Set job CPU nice value",
319		.minval	= -19,
320		.maxval	= 20,
321		.def	= "0",
322	},
323#ifdef FIO_HAVE_IOPRIO
324	{
325		.name	= "prio",
326		.type	= FIO_OPT_INT,
327		.cb	= str_prio_cb,
328		.help	= "Set job IO priority value",
329		.minval	= 0,
330		.maxval	= 7,
331	},
332	{
333		.name	= "prioclass",
334		.type	= FIO_OPT_INT,
335		.cb	= str_prioclass_cb,
336		.help	= "Set job IO priority class",
337		.minval	= 0,
338		.maxval	= 3,
339	},
340#endif
341	{
342		.name	= "thinktime",
343		.type	= FIO_OPT_INT,
344		.off1	= td_var_offset(thinktime),
345		.help	= "Idle time between IO buffers (usec)",
346		.def	= "0",
347	},
348	{
349		.name	= "thinktime_spin",
350		.type	= FIO_OPT_INT,
351		.off1	= td_var_offset(thinktime_spin),
352		.help	= "Start thinktime by spinning this amount (usec)",
353		.def	= "0",
354	},
355	{
356		.name	= "thinktime_blocks",
357		.type	= FIO_OPT_INT,
358		.off1	= td_var_offset(thinktime_blocks),
359		.help	= "IO buffer period between 'thinktime'",
360		.def	= "1",
361	},
362	{
363		.name	= "rate",
364		.type	= FIO_OPT_INT,
365		.off1	= td_var_offset(rate),
366		.help	= "Set bandwidth rate",
367	},
368	{
369		.name	= "ratemin",
370		.type	= FIO_OPT_INT,
371		.off1	= td_var_offset(ratemin),
372		.help	= "The bottom limit accepted",
373	},
374	{
375		.name	= "ratecycle",
376		.type	= FIO_OPT_INT,
377		.off1	= td_var_offset(ratecycle),
378		.help	= "Window average for rate limits (msec)",
379		.def	= "1000",
380	},
381	{
382		.name	= "invalidate",
383		.type	= FIO_OPT_BOOL,
384		.off1	= td_var_offset(invalidate_cache),
385		.help	= "Invalidate buffer/page cache prior to running job",
386		.def	= "1",
387	},
388	{
389		.name	= "sync",
390		.type	= FIO_OPT_BOOL,
391		.off1	= td_var_offset(sync_io),
392		.help	= "Use O_SYNC for buffered writes",
393		.def	= "0",
394	},
395	{
396		.name	= "bwavgtime",
397		.type	= FIO_OPT_INT,
398		.off1	= td_var_offset(bw_avg_time),
399		.help	= "Time window over which to calculate bandwidth (msec)",
400		.def	= "500",
401	},
402	{
403		.name	= "create_serialize",
404		.type	= FIO_OPT_BOOL,
405		.off1	= td_var_offset(create_serialize),
406		.help	= "Serialize creating of job files",
407		.def	= "1",
408	},
409	{
410		.name	= "create_fsync",
411		.type	= FIO_OPT_BOOL,
412		.off1	= td_var_offset(create_fsync),
413		.help	= "Fsync file after creation",
414		.def	= "1",
415	},
416	{
417		.name	= "cpuload",
418		.type	= FIO_OPT_INT,
419		.off1	= td_var_offset(cpuload),
420		.help	= "Use this percentage of CPU",
421	},
422	{
423		.name	= "cpuchunks",
424		.type	= FIO_OPT_INT,
425		.off1	= td_var_offset(cpucycle),
426		.help	= "Length of the CPU burn cycles",
427	},
428#ifdef FIO_HAVE_CPU_AFFINITY
429	{
430		.name	= "cpumask",
431		.type	= FIO_OPT_INT,
432		.cb	= str_cpumask_cb,
433		.help	= "CPU affinity mask",
434	},
435#endif
436	{
437		.name	= "end_fsync",
438		.type	= FIO_OPT_BOOL,
439		.off1	= td_var_offset(end_fsync),
440		.help	= "Include fsync at the end of job",
441		.def	= "0",
442	},
443	{
444		.name	= "unlink",
445		.type	= FIO_OPT_BOOL,
446		.off1	= td_var_offset(unlink),
447		.help	= "Unlink created files after job has completed",
448		.def	= "0",
449	},
450	{
451		.name	= "exitall",
452		.type	= FIO_OPT_STR_SET,
453		.cb	= str_exitall_cb,
454		.help	= "Terminate all jobs when one exits",
455	},
456	{
457		.name	= "stonewall",
458		.type	= FIO_OPT_STR_SET,
459		.off1	= td_var_offset(stonewall),
460		.help	= "Insert a hard barrier between this job and previous",
461	},
462	{
463		.name	= "thread",
464		.type	= FIO_OPT_STR_SET,
465		.off1	= td_var_offset(use_thread),
466		.help	= "Use threads instead of forks",
467	},
468	{
469		.name	= "write_bw_log",
470		.type	= FIO_OPT_STR_SET,
471		.off1	= td_var_offset(write_bw_log),
472		.help	= "Write log of bandwidth during run",
473	},
474	{
475		.name	= "write_lat_log",
476		.type	= FIO_OPT_STR_SET,
477		.off1	= td_var_offset(write_lat_log),
478		.help	= "Write log of latency during run",
479	},
480	{
481		.name	= "hugepage-size",
482		.type	= FIO_OPT_STR_VAL,
483		.off1	= td_var_offset(hugepage_size),
484		.help	= "When using hugepages, specify size of each page",
485		.def	= __stringify(FIO_HUGE_PAGE),
486	},
487	{
488		.name = NULL,
489	},
490};
491
492#define FIO_JOB_OPTS	(sizeof(options) / sizeof(struct fio_option))
493#define FIO_CMD_OPTS	(16)
494#define FIO_GETOPT_JOB	(0x89988998)
495
496/*
497 * Command line options. These will contain the above, plus a few
498 * extra that only pertain to fio itself and not jobs.
499 */
500static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
501	{
502		.name		= "output",
503		.has_arg	= required_argument,
504		.val		= 'o',
505	},
506	{
507		.name		= "timeout",
508		.has_arg	= required_argument,
509		.val		= 't',
510	},
511	{
512		.name		= "latency-log",
513		.has_arg	= required_argument,
514		.val		= 'l',
515	},
516	{
517		.name		= "bandwidth-log",
518		.has_arg	= required_argument,
519		.val		= 'b',
520	},
521	{
522		.name		= "minimal",
523		.has_arg	= optional_argument,
524		.val		= 'm',
525	},
526	{
527		.name		= "version",
528		.has_arg	= no_argument,
529		.val		= 'v',
530	},
531	{
532		.name		= "help",
533		.has_arg	= no_argument,
534		.val		= 'h',
535	},
536	{
537		.name		= "cmdhelp",
538		.has_arg	= required_argument,
539		.val		= 'c',
540	},
541	{
542		.name		= NULL,
543	},
544};
545
546static int def_timeout = 0;
547
548static char fio_version_string[] = "fio 1.11";
549
550static char **ini_file;
551static int max_jobs = MAX_JOBS;
552
553struct thread_data def_thread;
554struct thread_data *threads = NULL;
555
556int exitall_on_terminate = 0;
557int terse_output = 0;
558unsigned long long mlock_size = 0;
559FILE *f_out = NULL;
560FILE *f_err = NULL;
561
562static int write_lat_log = 0;
563int write_bw_log = 0;
564
565/*
566 * Return a free job structure.
567 */
568static struct thread_data *get_new_job(int global, struct thread_data *parent)
569{
570	struct thread_data *td;
571
572	if (global)
573		return &def_thread;
574	if (thread_number >= max_jobs)
575		return NULL;
576
577	td = &threads[thread_number++];
578	*td = *parent;
579
580	td->thread_number = thread_number;
581	return td;
582}
583
584static void put_job(struct thread_data *td)
585{
586	if (td == &def_thread)
587		return;
588
589	if (td->error)
590		fprintf(f_out, "fio: %s\n", td->verror);
591
592	memset(&threads[td->thread_number - 1], 0, sizeof(*td));
593	thread_number--;
594}
595
596/*
597 * Lazy way of fixing up options that depend on each other. We could also
598 * define option callback handlers, but this is easier.
599 */
600static void fixup_options(struct thread_data *td)
601{
602	if (!td->rwmixread && td->rwmixwrite)
603		td->rwmixread = 100 - td->rwmixwrite;
604
605	if (td->write_iolog_file && td->read_iolog_file) {
606		log_err("fio: read iolog overrides write_iolog\n");
607		free(td->write_iolog_file);
608		td->write_iolog_file = NULL;
609	}
610
611	if (td->io_ops->flags & FIO_SYNCIO)
612		td->iodepth = 1;
613	else {
614		if (!td->iodepth)
615			td->iodepth = td->nr_files;
616	}
617
618	/*
619	 * only really works for sequential io for now, and with 1 file
620	 */
621	if (td->zone_size && td_random(td) && td->nr_files == 1)
622		td->zone_size = 0;
623
624	/*
625	 * Reads can do overwrites, we always need to pre-create the file
626	 */
627	if (td_read(td) || td_rw(td))
628		td->overwrite = 1;
629
630	if (!td->min_bs[DDIR_READ])
631		td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
632	if (!td->max_bs[DDIR_READ])
633		td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
634	if (!td->min_bs[DDIR_WRITE])
635		td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
636	if (!td->max_bs[DDIR_WRITE])
637		td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
638
639	td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
640
641	if (td_read(td) && !td_rw(td))
642		td->verify = 0;
643
644	if (td->norandommap && td->verify != VERIFY_NONE) {
645		log_err("fio: norandommap given, verify disabled\n");
646		td->verify = VERIFY_NONE;
647	}
648	if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
649		log_err("fio: bs_unaligned may not work with raw io\n");
650
651	/*
652	 * O_DIRECT and char doesn't mix, clear that flag if necessary.
653	 */
654	if (td->filetype == FIO_TYPE_CHAR && td->odirect)
655		td->odirect = 0;
656
657	/*
658	 * thinktime_spin must be less than thinktime
659	 */
660	if (td->thinktime_spin > td->thinktime)
661		td->thinktime_spin = td->thinktime;
662
663	/*
664	 * The low water mark cannot be bigger than the iodepth
665	 */
666	if (td->iodepth_low > td->iodepth || !td->iodepth_low)
667		td->iodepth_low = td->iodepth;
668}
669
670/*
671 * This function leaks the buffer
672 */
673static char *to_kmg(unsigned int val)
674{
675	char *buf = malloc(32);
676	char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
677	char *p = post;
678
679	do {
680		if (val & 1023)
681			break;
682
683		val >>= 10;
684		p++;
685	} while (*p);
686
687	snprintf(buf, 31, "%u%c", val, *p);
688	return buf;
689}
690
691/*
692 * Adds a job to the list of things todo. Sanitizes the various options
693 * to make sure we don't have conflicts, and initializes various
694 * members of td.
695 */
696static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
697{
698	const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
699				   "randread", "randwrite", "randrw" };
700	struct stat sb;
701	int numjobs, i;
702	struct fio_file *f;
703
704	/*
705	 * the def_thread is just for options, it's not a real job
706	 */
707	if (td == &def_thread)
708		return 0;
709
710	assert(td->io_ops);
711
712	if (td->odirect)
713		td->io_ops->flags |= FIO_RAWIO;
714
715	td->filetype = FIO_TYPE_FILE;
716	if (td->filename && !lstat(td->filename, &sb)) {
717		if (S_ISBLK(sb.st_mode))
718			td->filetype = FIO_TYPE_BD;
719		else if (S_ISCHR(sb.st_mode))
720			td->filetype = FIO_TYPE_CHAR;
721	}
722
723	fixup_options(td);
724
725	if (td->filename)
726		td->nr_uniq_files = 1;
727	else
728		td->nr_uniq_files = td->nr_files;
729
730	if (td->filetype == FIO_TYPE_FILE || td->filename) {
731		char tmp[PATH_MAX];
732		int len = 0;
733
734		if (td->directory && td->directory[0] != '\0') {
735			if (lstat(td->directory, &sb) < 0) {
736				log_err("fio: %s is not a directory\n", td->directory);
737				td_verror(td, errno, "lstat");
738				return 1;
739			}
740			if (!S_ISDIR(sb.st_mode)) {
741				log_err("fio: %s is not a directory\n", td->directory);
742				return 1;
743			}
744			len = sprintf(tmp, "%s/", td->directory);
745		}
746
747		td->files = malloc(sizeof(struct fio_file) * td->nr_files);
748
749		for_each_file(td, f, i) {
750			memset(f, 0, sizeof(*f));
751			f->fd = -1;
752
753			if (td->filename)
754				sprintf(tmp + len, "%s", td->filename);
755			else
756				sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
757			f->file_name = strdup(tmp);
758		}
759	} else {
760		td->nr_files = 1;
761		td->files = malloc(sizeof(struct fio_file));
762		f = &td->files[0];
763
764		memset(f, 0, sizeof(*f));
765		f->fd = -1;
766		f->file_name = strdup(jobname);
767	}
768
769	for_each_file(td, f, i) {
770		f->file_size = td->total_file_size / td->nr_files;
771		f->file_offset = td->start_offset;
772	}
773
774	fio_sem_init(&td->mutex, 0);
775
776	td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
777	td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
778	td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
779
780	if (td->stonewall && td->thread_number > 1)
781		groupid++;
782
783	td->groupid = groupid;
784
785	if (setup_rate(td))
786		goto err;
787
788	if (td->write_lat_log) {
789		setup_log(&td->ts.slat_log);
790		setup_log(&td->ts.clat_log);
791	}
792	if (td->write_bw_log)
793		setup_log(&td->ts.bw_log);
794
795	if (!td->name)
796		td->name = strdup(jobname);
797
798	if (!terse_output) {
799		if (!job_add_num) {
800			if (td->io_ops->flags & FIO_CPUIO)
801				fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
802			else {
803				char *c1, *c2, *c3, *c4;
804
805				c1 = to_kmg(td->min_bs[DDIR_READ]);
806				c2 = to_kmg(td->max_bs[DDIR_READ]);
807				c3 = to_kmg(td->min_bs[DDIR_WRITE]);
808				c4 = to_kmg(td->max_bs[DDIR_WRITE]);
809
810				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);
811
812				free(c1);
813				free(c2);
814				free(c3);
815				free(c4);
816			}
817		} else if (job_add_num == 1)
818			fprintf(f_out, "...\n");
819	}
820
821	/*
822	 * recurse add identical jobs, clear numjobs and stonewall options
823	 * as they don't apply to sub-jobs
824	 */
825	numjobs = td->numjobs;
826	while (--numjobs) {
827		struct thread_data *td_new = get_new_job(0, td);
828
829		if (!td_new)
830			goto err;
831
832		td_new->numjobs = 1;
833		td_new->stonewall = 0;
834		job_add_num = numjobs - 1;
835
836		if (add_job(td_new, jobname, job_add_num))
837			goto err;
838	}
839	return 0;
840err:
841	put_job(td);
842	return -1;
843}
844
845/*
846 * Initialize the various random states we need (random io, block size ranges,
847 * read/write mix, etc).
848 */
849int init_random_state(struct thread_data *td)
850{
851	unsigned long seeds[5];
852	int fd, num_maps, blocks, i;
853	struct fio_file *f;
854
855	if (td->io_ops->flags & FIO_CPUIO)
856		return 0;
857
858	fd = open("/dev/urandom", O_RDONLY);
859	if (fd == -1) {
860		td_verror(td, errno, "open");
861		return 1;
862	}
863
864	if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
865		td_verror(td, EIO, "read");
866		close(fd);
867		return 1;
868	}
869
870	close(fd);
871
872	os_random_seed(seeds[0], &td->bsrange_state);
873	os_random_seed(seeds[1], &td->verify_state);
874	os_random_seed(seeds[2], &td->rwmix_state);
875
876	if (td->file_service_type == FIO_FSERVICE_RANDOM)
877		os_random_seed(seeds[3], &td->next_file_state);
878
879	if (!td_random(td))
880		return 0;
881
882	if (td->rand_repeatable)
883		seeds[4] = FIO_RANDSEED * td->thread_number;
884
885	if (!td->norandommap) {
886		for_each_file(td, f, i) {
887			blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
888			num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
889			f->file_map = malloc(num_maps * sizeof(long));
890			f->num_maps = num_maps;
891			memset(f->file_map, 0, num_maps * sizeof(long));
892		}
893	}
894
895	os_random_seed(seeds[4], &td->random_state);
896	return 0;
897}
898
899static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
900{
901#ifdef FIO_HAVE_CPU_AFFINITY
902	unsigned int i;
903
904	CPU_ZERO(&cpumask);
905
906	for (i = 0; i < sizeof(int) * 8; i++) {
907		if ((1 << i) & cpu)
908			CPU_SET(i, &cpumask);
909	}
910#endif
911}
912
913static int is_empty_or_comment(char *line)
914{
915	unsigned int i;
916
917	for (i = 0; i < strlen(line); i++) {
918		if (line[i] == ';')
919			return 1;
920		if (line[i] == '#')
921			return 1;
922		if (!isspace(line[i]) && !iscntrl(line[i]))
923			return 0;
924	}
925
926	return 1;
927}
928
929static int str_rw_cb(void *data, const char *mem)
930{
931	struct thread_data *td = data;
932
933	if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
934		td->td_ddir = TD_DDIR_READ;
935		return 0;
936	} else if (!strncmp(mem, "randread", 8)) {
937		td->td_ddir = TD_DDIR_READ | TD_DDIR_RAND;
938		return 0;
939	} else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
940		td->td_ddir = TD_DDIR_WRITE;
941		return 0;
942	} else if (!strncmp(mem, "randwrite", 9)) {
943		td->td_ddir = TD_DDIR_WRITE | TD_DDIR_RAND;
944		return 0;
945	} else if (!strncmp(mem, "rw", 2)) {
946		td->td_ddir = TD_DDIR_RW;
947		return 0;
948	} else if (!strncmp(mem, "randrw", 6)) {
949		td->td_ddir = TD_DDIR_RW | TD_DDIR_RAND;
950		return 0;
951	}
952
953	log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
954	return 1;
955}
956
957static int str_verify_cb(void *data, const char *mem)
958{
959	struct thread_data *td = data;
960
961	if (!strncmp(mem, "0", 1)) {
962		td->verify = VERIFY_NONE;
963		return 0;
964	} else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
965		td->verify = VERIFY_MD5;
966		return 0;
967	} else if (!strncmp(mem, "crc32", 5)) {
968		td->verify = VERIFY_CRC32;
969		return 0;
970	}
971
972	log_err("fio: verify types: md5, crc32\n");
973	return 1;
974}
975
976/*
977 * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
978 */
979static char *get_mmap_file(const char *str)
980{
981	char *p = strstr(str, ":");
982
983	if (!p)
984		return NULL;
985
986	p++;
987	strip_blank_front(&p);
988	strip_blank_end(p);
989	return strdup(p);
990}
991
992static int str_mem_cb(void *data, const char *mem)
993{
994	struct thread_data *td = data;
995
996	if (!strncmp(mem, "malloc", 6)) {
997		td->mem_type = MEM_MALLOC;
998		return 0;
999	} else if (!strncmp(mem, "mmaphuge", 8)) {
1000#ifdef FIO_HAVE_HUGETLB
1001		/*
1002		 * mmaphuge must be appended with the actual file
1003		 */
1004		td->mmapfile = get_mmap_file(mem);
1005		if (!td->mmapfile) {
1006			log_err("fio: mmaphuge:/path/to/file\n");
1007			return 1;
1008		}
1009
1010		td->mem_type = MEM_MMAPHUGE;
1011		return 0;
1012#else
1013		log_err("fio: mmaphuge not available\n");
1014		return 1;
1015#endif
1016	} else if (!strncmp(mem, "mmap", 4)) {
1017		/*
1018		 * Check if the user wants file backed memory. It's ok
1019		 * if there's no file given, we'll just use anon mamp then.
1020		 */
1021		td->mmapfile = get_mmap_file(mem);
1022		td->mem_type = MEM_MMAP;
1023		return 0;
1024	} else if (!strncmp(mem, "shmhuge", 7)) {
1025#ifdef FIO_HAVE_HUGETLB
1026		td->mem_type = MEM_SHMHUGE;
1027		return 0;
1028#else
1029		log_err("fio: shmhuge not available\n");
1030		return 1;
1031#endif
1032	} else if (!strncmp(mem, "shm", 3)) {
1033		td->mem_type = MEM_SHM;
1034		return 0;
1035	}
1036
1037	log_err("fio: mem type: malloc, shm, shmhuge, mmap, mmaphuge\n");
1038	return 1;
1039}
1040
1041static int str_ioengine_cb(void *data, const char *str)
1042{
1043	struct thread_data *td = data;
1044
1045	td->io_ops = load_ioengine(td, str);
1046	if (td->io_ops)
1047		return 0;
1048
1049	log_err("fio: ioengine= libaio, posixaio, sync, syslet-rw, mmap, sgio, splice, cpu, null\n");
1050	log_err("fio: or specify path to dynamic ioengine module\n");
1051	return 1;
1052}
1053
1054static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1055{
1056	mlock_size = *val;
1057	return 0;
1058}
1059
1060#ifdef FIO_HAVE_IOPRIO
1061static int str_prioclass_cb(void *data, unsigned int *val)
1062{
1063	struct thread_data *td = data;
1064
1065	td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1066	return 0;
1067}
1068
1069static int str_prio_cb(void *data, unsigned int *val)
1070{
1071	struct thread_data *td = data;
1072
1073	td->ioprio |= *val;
1074	return 0;
1075}
1076#endif
1077
1078static int str_exitall_cb(void)
1079{
1080	exitall_on_terminate = 1;
1081	return 0;
1082}
1083
1084static int str_cpumask_cb(void *data, unsigned int *val)
1085{
1086	struct thread_data *td = data;
1087
1088	fill_cpu_mask(td->cpumask, *val);
1089	return 0;
1090}
1091
1092static int str_file_service_cb(void *data, const char *str)
1093{
1094	struct thread_data *td = data;
1095
1096	if (!strncmp(str, "random", 6)) {
1097		td->file_service_type = FIO_FSERVICE_RANDOM;
1098		return 0;
1099	} else if (!strncmp(str, "roundrobin", 10)) {
1100		td->file_service_type = FIO_FSERVICE_RR;
1101		return 0;
1102	}
1103
1104	log_err("fio: file_service= random, roundrobin\n");
1105	return 1;
1106}
1107
1108/*
1109 * This is our [ini] type file parser.
1110 */
1111static int parse_jobs_ini(char *file, int stonewall_flag)
1112{
1113	unsigned int global;
1114	struct thread_data *td;
1115	char *string, *name;
1116	fpos_t off;
1117	FILE *f;
1118	char *p;
1119	int ret = 0, stonewall;
1120
1121	f = fopen(file, "r");
1122	if (!f) {
1123		perror("fopen job file");
1124		return 1;
1125	}
1126
1127	string = malloc(4096);
1128	name = malloc(256);
1129	memset(name, 0, 256);
1130
1131	stonewall = stonewall_flag;
1132	do {
1133		p = fgets(string, 4095, f);
1134		if (!p)
1135			break;
1136		if (is_empty_or_comment(p))
1137			continue;
1138		if (sscanf(p, "[%255s]", name) != 1)
1139			continue;
1140
1141		global = !strncmp(name, "global", 6);
1142
1143		name[strlen(name) - 1] = '\0';
1144
1145		td = get_new_job(global, &def_thread);
1146		if (!td) {
1147			ret = 1;
1148			break;
1149		}
1150
1151		/*
1152		 * Seperate multiple job files by a stonewall
1153		 */
1154		if (!global && stonewall) {
1155			td->stonewall = stonewall;
1156			stonewall = 0;
1157		}
1158
1159		fgetpos(f, &off);
1160		while ((p = fgets(string, 4096, f)) != NULL) {
1161			if (is_empty_or_comment(p))
1162				continue;
1163
1164			strip_blank_front(&p);
1165
1166			if (p[0] == '[')
1167				break;
1168
1169			strip_blank_end(p);
1170
1171			fgetpos(f, &off);
1172
1173			/*
1174			 * Don't break here, continue parsing options so we
1175			 * dump all the bad ones. Makes trial/error fixups
1176			 * easier on the user.
1177			 */
1178			ret |= parse_option(p, options, td);
1179		}
1180
1181		if (!ret) {
1182			fsetpos(f, &off);
1183			ret = add_job(td, name, 0);
1184		} else {
1185			log_err("fio: job %s dropped\n", name);
1186			put_job(td);
1187		}
1188	} while (!ret);
1189
1190	free(string);
1191	free(name);
1192	fclose(f);
1193	return ret;
1194}
1195
1196static int fill_def_thread(void)
1197{
1198	memset(&def_thread, 0, sizeof(def_thread));
1199
1200	if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1201		perror("sched_getaffinity");
1202		return 1;
1203	}
1204
1205	/*
1206	 * fill default options
1207	 */
1208	fill_default_options(&def_thread, options);
1209
1210	def_thread.timeout = def_timeout;
1211	def_thread.write_bw_log = write_bw_log;
1212	def_thread.write_lat_log = write_lat_log;
1213
1214#ifdef FIO_HAVE_DISK_UTIL
1215	def_thread.do_disk_util = 1;
1216#endif
1217
1218	return 0;
1219}
1220
1221static void usage(void)
1222{
1223	printf("%s\n", fio_version_string);
1224	printf("\t--output\tWrite output to file\n");
1225	printf("\t--timeout\tRuntime in seconds\n");
1226	printf("\t--latency-log\tGenerate per-job latency logs\n");
1227	printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1228	printf("\t--minimal\tMinimal (terse) output\n");
1229	printf("\t--version\tPrint version info and exit\n");
1230	printf("\t--help\t\tPrint this page\n");
1231	printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1232}
1233
1234static int parse_cmd_line(int argc, char *argv[])
1235{
1236	struct thread_data *td = NULL;
1237	int c, ini_idx = 0, lidx, ret;
1238
1239	while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1240		switch (c) {
1241		case 't':
1242			def_timeout = atoi(optarg);
1243			break;
1244		case 'l':
1245			write_lat_log = 1;
1246			break;
1247		case 'w':
1248			write_bw_log = 1;
1249			break;
1250		case 'o':
1251			f_out = fopen(optarg, "w+");
1252			if (!f_out) {
1253				perror("fopen output");
1254				exit(1);
1255			}
1256			f_err = f_out;
1257			break;
1258		case 'm':
1259			terse_output = 1;
1260			break;
1261		case 'h':
1262			usage();
1263			exit(0);
1264		case 'c':
1265			ret = show_cmd_help(options, optarg);
1266			exit(ret);
1267		case 'v':
1268			printf("%s\n", fio_version_string);
1269			exit(0);
1270		case FIO_GETOPT_JOB: {
1271			const char *opt = long_options[lidx].name;
1272			char *val = optarg;
1273
1274			if (!strncmp(opt, "name", 4) && td) {
1275				ret = add_job(td, td->name ?: "fio", 0);
1276				if (ret) {
1277					put_job(td);
1278					return 0;
1279				}
1280				td = NULL;
1281			}
1282			if (!td) {
1283				int global = !strncmp(val, "global", 6);
1284
1285				td = get_new_job(global, &def_thread);
1286				if (!td)
1287					return 0;
1288			}
1289
1290			ret = parse_cmd_option(opt, val, options, td);
1291			if (ret) {
1292				log_err("fio: job dropped\n");
1293				put_job(td);
1294				td = NULL;
1295			}
1296			break;
1297		}
1298		default:
1299			break;
1300		}
1301	}
1302
1303	if (td) {
1304		ret = add_job(td, td->name ?: "fio", 0);
1305		if (ret)
1306			put_job(td);
1307	}
1308
1309	while (optind < argc) {
1310		ini_idx++;
1311		ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1312		ini_file[ini_idx - 1] = strdup(argv[optind]);
1313		optind++;
1314	}
1315
1316	return ini_idx;
1317}
1318
1319static void free_shm(void)
1320{
1321	struct shmid_ds sbuf;
1322
1323	if (threads) {
1324		shmdt((void *) threads);
1325		threads = NULL;
1326		shmctl(shm_id, IPC_RMID, &sbuf);
1327	}
1328}
1329
1330/*
1331 * The thread area is shared between the main process and the job
1332 * threads/processes. So setup a shared memory segment that will hold
1333 * all the job info.
1334 */
1335static int setup_thread_area(void)
1336{
1337	/*
1338	 * 1024 is too much on some machines, scale max_jobs if
1339	 * we get a failure that looks like too large a shm segment
1340	 */
1341	do {
1342		size_t size = max_jobs * sizeof(struct thread_data);
1343
1344		shm_id = shmget(0, size, IPC_CREAT | 0600);
1345		if (shm_id != -1)
1346			break;
1347		if (errno != EINVAL) {
1348			perror("shmget");
1349			break;
1350		}
1351
1352		max_jobs >>= 1;
1353	} while (max_jobs);
1354
1355	if (shm_id == -1)
1356		return 1;
1357
1358	threads = shmat(shm_id, NULL, 0);
1359	if (threads == (void *) -1) {
1360		perror("shmat");
1361		return 1;
1362	}
1363
1364	atexit(free_shm);
1365	return 0;
1366}
1367
1368/*
1369 * Copy the fio options into the long options map, so we mirror
1370 * job and cmd line options.
1371 */
1372static void dupe_job_options(void)
1373{
1374	struct fio_option *o;
1375	unsigned int i;
1376
1377	i = 0;
1378	while (long_options[i].name)
1379		i++;
1380
1381	o = &options[0];
1382	while (o->name) {
1383		long_options[i].name = o->name;
1384		long_options[i].val = FIO_GETOPT_JOB;
1385		if (o->type == FIO_OPT_STR_SET)
1386			long_options[i].has_arg = no_argument;
1387		else
1388			long_options[i].has_arg = required_argument;
1389
1390		i++;
1391		o++;
1392		assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1393	}
1394}
1395
1396int parse_options(int argc, char *argv[])
1397{
1398	int job_files, i;
1399
1400	f_out = stdout;
1401	f_err = stderr;
1402
1403	options_init(options);
1404
1405	dupe_job_options();
1406
1407	if (setup_thread_area())
1408		return 1;
1409	if (fill_def_thread())
1410		return 1;
1411
1412	job_files = parse_cmd_line(argc, argv);
1413
1414	for (i = 0; i < job_files; i++) {
1415		if (fill_def_thread())
1416			return 1;
1417		if (parse_jobs_ini(ini_file[i], i))
1418			return 1;
1419		free(ini_file[i]);
1420	}
1421
1422	free(ini_file);
1423
1424	if (!thread_number) {
1425		log_err("No jobs defined(s)\n");
1426		return 1;
1427	}
1428
1429	return 0;
1430}
1431