c8f27ade76e7f2aafaef4e98320d189c2a26e347
[fio.git] / init.c
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
25 static int str_rw_cb(void *, const char *);
26 static int str_ioengine_cb(void *, const char *);
27 static int str_mem_cb(void *, const char *);
28 static int str_verify_cb(void *, const char *);
29 static int str_lockmem_cb(void *, unsigned long *);
30 #ifdef FIO_HAVE_IOPRIO
31 static int str_prio_cb(void *, unsigned int *);
32 static int str_prioclass_cb(void *, unsigned int *);
33 #endif
34 static int str_exitall_cb(void);
35 static 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  */
43 static 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", "net", "syslet-rw" },
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   = "iodepth_low",
95                 .type   = FIO_OPT_INT,
96                 .off1   = td_var_offset(iodepth_low),
97                 .help   = "Low water mark for queuing depth",
98         },
99         {
100                 .name   = "size",
101                 .type   = FIO_OPT_STR_VAL,
102                 .off1   = td_var_offset(total_file_size),
103                 .help   = "Size of device or file",
104         },
105         {
106                 .name   = "bs",
107                 .type   = FIO_OPT_STR_VAL_INT,
108                 .off1   = td_var_offset(bs[DDIR_READ]),
109                 .off2   = td_var_offset(bs[DDIR_WRITE]),
110                 .help   = "Block size unit",
111                 .def    = "4k",
112         },
113         {
114                 .name   = "bsrange",
115                 .type   = FIO_OPT_RANGE,
116                 .off1   = td_var_offset(min_bs[DDIR_READ]),
117                 .off2   = td_var_offset(max_bs[DDIR_READ]),
118                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
119                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
120                 .help   = "Set block size range (in more detail than bs)",
121         },
122         {
123                 .name   = "bs_unaligned",
124                 .type   = FIO_OPT_STR_SET,
125                 .off1   = td_var_offset(bs_unaligned),
126                 .help   = "Don't sector align IO buffer sizes",
127         },
128         {
129                 .name   = "offset",
130                 .type   = FIO_OPT_STR_VAL,
131                 .off1   = td_var_offset(start_offset),
132                 .help   = "Start IO from this offset",
133                 .def    = "0",
134         },
135         {
136                 .name   = "randrepeat",
137                 .type   = FIO_OPT_BOOL,
138                 .off1   = td_var_offset(rand_repeatable),
139                 .help   = "Use repeatable random IO pattern",
140                 .def    = "1",
141         },
142         {
143                 .name   = "norandommap",
144                 .type   = FIO_OPT_STR_SET,
145                 .off1   = td_var_offset(norandommap),
146                 .help   = "Accept potential duplicate random blocks",
147         },
148         {
149                 .name   = "nrfiles",
150                 .type   = FIO_OPT_INT,
151                 .off1   = td_var_offset(nr_files),
152                 .help   = "Split job workload between this number of files",
153                 .def    = "1",
154         },
155         {
156                 .name   = "fsync",
157                 .type   = FIO_OPT_INT,
158                 .off1   = td_var_offset(fsync_blocks),
159                 .help   = "Issue fsync for writes every given number of blocks",
160                 .def    = "0",
161         },
162         {
163                 .name   = "direct",
164                 .type   = FIO_OPT_BOOL,
165                 .off1   = td_var_offset(odirect),
166                 .help   = "Use O_DIRECT IO (negates buffered)",
167                 .def    = "0",
168         },
169         {
170                 .name   = "buffered",
171                 .type   = FIO_OPT_BOOL,
172                 .off1   = td_var_offset(odirect),
173                 .neg    = 1,
174                 .help   = "Use buffered IO (negates direct)",
175                 .def    = "1",
176         },
177         {
178                 .name   = "overwrite",
179                 .type   = FIO_OPT_BOOL,
180                 .off1   = td_var_offset(overwrite),
181                 .help   = "When writing, set whether to overwrite current data",
182                 .def    = "0",
183         },
184         {
185                 .name   = "loops",
186                 .type   = FIO_OPT_INT,
187                 .off1   = td_var_offset(loops),
188                 .help   = "Number of times to run the job",
189                 .def    = "1",
190         },
191         {
192                 .name   = "numjobs",
193                 .type   = FIO_OPT_INT,
194                 .off1   = td_var_offset(numjobs),
195                 .help   = "Duplicate this job this many times",
196                 .def    = "1",
197         },
198         {
199                 .name   = "startdelay",
200                 .type   = FIO_OPT_INT,
201                 .off1   = td_var_offset(start_delay),
202                 .help   = "Only start job when this period has passed",
203                 .def    = "0",
204         },
205         {
206                 .name   = "runtime",
207                 .alias  = "timeout",
208                 .type   = FIO_OPT_STR_VAL_TIME,
209                 .off1   = td_var_offset(timeout),
210                 .help   = "Stop workload when this amount of time has passed",
211                 .def    = "0",
212         },
213         {
214                 .name   = "mem",
215                 .type   = FIO_OPT_STR,
216                 .cb     = str_mem_cb,
217                 .help   = "Backing type for IO buffers",
218                 .def    = "malloc",
219                 .posval =  { "malloc", "shm", "shmhuge", "mmap", "mmaphuge", },
220         },
221         {
222                 .name   = "verify",
223                 .type   = FIO_OPT_STR,
224                 .cb     = str_verify_cb,
225                 .help   = "Verify sum function",
226                 .def    = "0",
227                 .posval = { "crc32", "md5", },
228         },
229         {
230                 .name   = "write_iolog",
231                 .type   = FIO_OPT_STR_STORE,
232                 .off1   = td_var_offset(write_iolog_file),
233                 .help   = "Store IO pattern to file",
234         },
235         {
236                 .name   = "read_iolog",
237                 .type   = FIO_OPT_STR_STORE,
238                 .off1   = td_var_offset(read_iolog_file),
239                 .help   = "Playback IO pattern from file",
240         },
241         {
242                 .name   = "exec_prerun",
243                 .type   = FIO_OPT_STR_STORE,
244                 .off1   = td_var_offset(exec_prerun),
245                 .help   = "Execute this file prior to running job",
246         },
247         {
248                 .name   = "exec_postrun",
249                 .type   = FIO_OPT_STR_STORE,
250                 .off1   = td_var_offset(exec_postrun),
251                 .help   = "Execute this file after running job",
252         },
253 #ifdef FIO_HAVE_IOSCHED_SWITCH
254         {
255                 .name   = "ioscheduler",
256                 .type   = FIO_OPT_STR_STORE,
257                 .off1   = td_var_offset(ioscheduler),
258                 .help   = "Use this IO scheduler on the backing device",
259         },
260 #endif
261         {
262                 .name   = "zonesize",
263                 .type   = FIO_OPT_STR_VAL,
264                 .off1   = td_var_offset(zone_size),
265                 .help   = "Give size of an IO zone",
266                 .def    = "0",
267         },
268         {
269                 .name   = "zoneskip",
270                 .type   = FIO_OPT_STR_VAL,
271                 .off1   = td_var_offset(zone_skip),
272                 .help   = "Space between IO zones",
273                 .def    = "0",
274         },
275         {
276                 .name   = "lockmem",
277                 .type   = FIO_OPT_STR_VAL,
278                 .cb     = str_lockmem_cb,
279                 .help   = "Lock down this amount of memory",
280                 .def    = "0",
281         },
282         {
283                 .name   = "rwmixcycle",
284                 .type   = FIO_OPT_INT,
285                 .off1   = td_var_offset(rwmixcycle),
286                 .help   = "Cycle period for mixed read/write workloads (msec)",
287                 .def    = "500",
288         },
289         {
290                 .name   = "rwmixread",
291                 .type   = FIO_OPT_INT,
292                 .off1   = td_var_offset(rwmixread),
293                 .maxval = 100,
294                 .help   = "Percentage of mixed workload that is reads",
295                 .def    = "50",
296         },
297         {
298                 .name   = "rwmixwrite",
299                 .type   = FIO_OPT_INT,
300                 .off1   = td_var_offset(rwmixwrite),
301                 .maxval = 100,
302                 .help   = "Percentage of mixed workload that is writes",
303                 .def    = "50",
304         },
305         {
306                 .name   = "nice",
307                 .type   = FIO_OPT_INT,
308                 .off1   = td_var_offset(nice),
309                 .help   = "Set job CPU nice value",
310                 .minval = -19,
311                 .maxval = 20,
312                 .def    = "0",
313         },
314 #ifdef FIO_HAVE_IOPRIO
315         {
316                 .name   = "prio",
317                 .type   = FIO_OPT_INT,
318                 .cb     = str_prio_cb,
319                 .help   = "Set job IO priority value",
320                 .minval = 0,
321                 .maxval = 7,
322         },
323         {
324                 .name   = "prioclass",
325                 .type   = FIO_OPT_INT,
326                 .cb     = str_prioclass_cb,
327                 .help   = "Set job IO priority class",
328                 .minval = 0,
329                 .maxval = 3,
330         },
331 #endif
332         {
333                 .name   = "thinktime",
334                 .type   = FIO_OPT_INT,
335                 .off1   = td_var_offset(thinktime),
336                 .help   = "Idle time between IO buffers (usec)",
337                 .def    = "0",
338         },
339         {
340                 .name   = "thinktime_spin",
341                 .type   = FIO_OPT_INT,
342                 .off1   = td_var_offset(thinktime_spin),
343                 .help   = "Start thinktime by spinning this amount (usec)",
344                 .def    = "0",
345         },
346         {
347                 .name   = "thinktime_blocks",
348                 .type   = FIO_OPT_INT,
349                 .off1   = td_var_offset(thinktime_blocks),
350                 .help   = "IO buffer period between 'thinktime'",
351                 .def    = "1",
352         },
353         {
354                 .name   = "rate",
355                 .type   = FIO_OPT_INT,
356                 .off1   = td_var_offset(rate),
357                 .help   = "Set bandwidth rate",
358         },
359         {
360                 .name   = "ratemin",
361                 .type   = FIO_OPT_INT,
362                 .off1   = td_var_offset(ratemin),
363                 .help   = "The bottom limit accepted",
364         },
365         {
366                 .name   = "ratecycle",
367                 .type   = FIO_OPT_INT,
368                 .off1   = td_var_offset(ratecycle),
369                 .help   = "Window average for rate limits (msec)",
370                 .def    = "1000",
371         },
372         {
373                 .name   = "invalidate",
374                 .type   = FIO_OPT_BOOL,
375                 .off1   = td_var_offset(invalidate_cache),
376                 .help   = "Invalidate buffer/page cache prior to running job",
377                 .def    = "1",
378         },
379         {
380                 .name   = "sync",
381                 .type   = FIO_OPT_BOOL,
382                 .off1   = td_var_offset(sync_io),
383                 .help   = "Use O_SYNC for buffered writes",
384                 .def    = "0",
385         },
386         {
387                 .name   = "bwavgtime",
388                 .type   = FIO_OPT_INT,
389                 .off1   = td_var_offset(bw_avg_time),
390                 .help   = "Time window over which to calculate bandwidth (msec)",
391                 .def    = "500",
392         },
393         {
394                 .name   = "create_serialize",
395                 .type   = FIO_OPT_BOOL,
396                 .off1   = td_var_offset(create_serialize),
397                 .help   = "Serialize creating of job files",
398                 .def    = "1",
399         },
400         {
401                 .name   = "create_fsync",
402                 .type   = FIO_OPT_BOOL,
403                 .off1   = td_var_offset(create_fsync),
404                 .help   = "Fsync file after creation",
405                 .def    = "1",
406         },
407         {
408                 .name   = "cpuload",
409                 .type   = FIO_OPT_INT,
410                 .off1   = td_var_offset(cpuload),
411                 .help   = "Use this percentage of CPU",
412         },
413         {
414                 .name   = "cpuchunks",
415                 .type   = FIO_OPT_INT,
416                 .off1   = td_var_offset(cpucycle),
417                 .help   = "Length of the CPU burn cycles",
418         },
419 #ifdef FIO_HAVE_CPU_AFFINITY
420         {
421                 .name   = "cpumask",
422                 .type   = FIO_OPT_INT,
423                 .cb     = str_cpumask_cb,
424                 .help   = "CPU affinity mask",
425         },
426 #endif
427         {
428                 .name   = "end_fsync",
429                 .type   = FIO_OPT_BOOL,
430                 .off1   = td_var_offset(end_fsync),
431                 .help   = "Include fsync at the end of job",
432                 .def    = "0",
433         },
434         {
435                 .name   = "unlink",
436                 .type   = FIO_OPT_BOOL,
437                 .off1   = td_var_offset(unlink),
438                 .help   = "Unlink created files after job has completed",
439                 .def    = "0",
440         },
441         {
442                 .name   = "exitall",
443                 .type   = FIO_OPT_STR_SET,
444                 .cb     = str_exitall_cb,
445                 .help   = "Terminate all jobs when one exits",
446         },
447         {
448                 .name   = "stonewall",
449                 .type   = FIO_OPT_STR_SET,
450                 .off1   = td_var_offset(stonewall),
451                 .help   = "Insert a hard barrier between this job and previous",
452         },
453         {
454                 .name   = "thread",
455                 .type   = FIO_OPT_STR_SET,
456                 .off1   = td_var_offset(use_thread),
457                 .help   = "Use threads instead of forks",
458         },
459         {
460                 .name   = "write_bw_log",
461                 .type   = FIO_OPT_STR_SET,
462                 .off1   = td_var_offset(write_bw_log),
463                 .help   = "Write log of bandwidth during run",
464         },
465         {
466                 .name   = "write_lat_log",
467                 .type   = FIO_OPT_STR_SET,
468                 .off1   = td_var_offset(write_lat_log),
469                 .help   = "Write log of latency during run",
470         },
471         {
472                 .name   = "hugepage-size",
473                 .type   = FIO_OPT_STR_VAL,
474                 .off1   = td_var_offset(hugepage_size),
475                 .help   = "When using hugepages, specify size of each page",
476                 .def    = __stringify(FIO_HUGE_PAGE),
477         },
478         {
479                 .name = NULL,
480         },
481 };
482
483 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
484 #define FIO_CMD_OPTS    (16)
485 #define FIO_GETOPT_JOB  (0x89988998)
486
487 /*
488  * Command line options. These will contain the above, plus a few
489  * extra that only pertain to fio itself and not jobs.
490  */
491 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
492         {
493                 .name           = "output",
494                 .has_arg        = required_argument,
495                 .val            = 'o',
496         },
497         {
498                 .name           = "timeout",
499                 .has_arg        = required_argument,
500                 .val            = 't',
501         },
502         {
503                 .name           = "latency-log",
504                 .has_arg        = required_argument,
505                 .val            = 'l',
506         },
507         {
508                 .name           = "bandwidth-log",
509                 .has_arg        = required_argument,
510                 .val            = 'b',
511         },
512         {
513                 .name           = "minimal",
514                 .has_arg        = optional_argument,
515                 .val            = 'm',
516         },
517         {
518                 .name           = "version",
519                 .has_arg        = no_argument,
520                 .val            = 'v',
521         },
522         {
523                 .name           = "help",
524                 .has_arg        = no_argument,
525                 .val            = 'h',
526         },
527         {
528                 .name           = "cmdhelp",
529                 .has_arg        = required_argument,
530                 .val            = 'c',
531         },
532         {
533                 .name           = NULL,
534         },
535 };
536
537 static int def_timeout = 0;
538
539 static char fio_version_string[] = "fio 1.11";
540
541 static char **ini_file;
542 static int max_jobs = MAX_JOBS;
543
544 struct thread_data def_thread;
545 struct thread_data *threads = NULL;
546
547 int exitall_on_terminate = 0;
548 int terse_output = 0;
549 unsigned long long mlock_size = 0;
550 FILE *f_out = NULL;
551 FILE *f_err = NULL;
552
553 static int write_lat_log = 0;
554 int write_bw_log = 0;
555
556 /*
557  * Return a free job structure.
558  */
559 static struct thread_data *get_new_job(int global, struct thread_data *parent)
560 {
561         struct thread_data *td;
562
563         if (global)
564                 return &def_thread;
565         if (thread_number >= max_jobs)
566                 return NULL;
567
568         td = &threads[thread_number++];
569         *td = *parent;
570
571         td->thread_number = thread_number;
572         return td;
573 }
574
575 static void put_job(struct thread_data *td)
576 {
577         if (td == &def_thread)
578                 return;
579
580         if (td->error)
581                 fprintf(f_out, "fio: %s\n", td->verror);
582
583         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
584         thread_number--;
585 }
586
587 /*
588  * Lazy way of fixing up options that depend on each other. We could also
589  * define option callback handlers, but this is easier.
590  */
591 static void fixup_options(struct thread_data *td)
592 {
593         if (!td->rwmixread && td->rwmixwrite)
594                 td->rwmixread = 100 - td->rwmixwrite;
595
596         if (td->write_iolog_file && td->read_iolog_file) {
597                 log_err("fio: read iolog overrides write_iolog\n");
598                 free(td->write_iolog_file);
599                 td->write_iolog_file = NULL;
600         }
601
602         if (td->io_ops->flags & FIO_SYNCIO)
603                 td->iodepth = 1;
604         else {
605                 if (!td->iodepth)
606                         td->iodepth = td->nr_files;
607         }
608
609         /*
610          * only really works for sequential io for now, and with 1 file
611          */
612         if (td->zone_size && !td->sequential && td->nr_files == 1)
613                 td->zone_size = 0;
614
615         /*
616          * Reads can do overwrites, we always need to pre-create the file
617          */
618         if (td_read(td) || td_rw(td))
619                 td->overwrite = 1;
620
621         if (!td->min_bs[DDIR_READ])
622                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
623         if (!td->max_bs[DDIR_READ])
624                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
625         if (!td->min_bs[DDIR_WRITE])
626                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
627         if (!td->max_bs[DDIR_WRITE])
628                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
629
630         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
631
632         if (td_read(td) && !td_rw(td))
633                 td->verify = 0;
634
635         if (td->norandommap && td->verify != VERIFY_NONE) {
636                 log_err("fio: norandommap given, verify disabled\n");
637                 td->verify = VERIFY_NONE;
638         }
639         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
640                 log_err("fio: bs_unaligned may not work with raw io\n");
641
642         /*
643          * O_DIRECT and char doesn't mix, clear that flag if necessary.
644          */
645         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
646                 td->odirect = 0;
647
648         /*
649          * thinktime_spin must be less than thinktime
650          */
651         if (td->thinktime_spin > td->thinktime)
652                 td->thinktime_spin = td->thinktime;
653
654         /*
655          * The low water mark cannot be bigger than the iodepth
656          */
657         if (td->iodepth_low > td->iodepth || !td->iodepth_low)
658                 td->iodepth_low = td->iodepth;
659 }
660
661 /*
662  * This function leaks the buffer
663  */
664 static char *to_kmg(unsigned int val)
665 {
666         char *buf = malloc(32);
667         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
668         char *p = post;
669
670         do {
671                 if (val & 1023)
672                         break;
673
674                 val >>= 10;
675                 p++;
676         } while (*p);
677
678         snprintf(buf, 31, "%u%c", val, *p);
679         return buf;
680 }
681
682 /*
683  * Adds a job to the list of things todo. Sanitizes the various options
684  * to make sure we don't have conflicts, and initializes various
685  * members of td.
686  */
687 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
688 {
689         const char *ddir_str[] = { "read", "write", "randread", "randwrite",
690                                    "rw", NULL, "randrw" };
691         struct stat sb;
692         int numjobs, ddir, i;
693         struct fio_file *f;
694
695         /*
696          * the def_thread is just for options, it's not a real job
697          */
698         if (td == &def_thread)
699                 return 0;
700
701         assert(td->io_ops);
702
703         if (td->odirect)
704                 td->io_ops->flags |= FIO_RAWIO;
705
706         td->filetype = FIO_TYPE_FILE;
707         if (td->filename && !lstat(td->filename, &sb)) {
708                 if (S_ISBLK(sb.st_mode))
709                         td->filetype = FIO_TYPE_BD;
710                 else if (S_ISCHR(sb.st_mode))
711                         td->filetype = FIO_TYPE_CHAR;
712         }
713
714         fixup_options(td);
715
716         if (td->filename)
717                 td->nr_uniq_files = 1;
718         else
719                 td->nr_uniq_files = td->nr_files;
720
721         if (td->filetype == FIO_TYPE_FILE || td->filename) {
722                 char tmp[PATH_MAX];
723                 int len = 0;
724
725                 if (td->directory && td->directory[0] != '\0') {
726                         if (lstat(td->directory, &sb) < 0) {
727                                 log_err("fio: %s is not a directory\n", td->directory);
728                                 td_verror(td, errno, "lstat");
729                                 return 1;
730                         }
731                         if (!S_ISDIR(sb.st_mode)) {
732                                 log_err("fio: %s is not a directory\n", td->directory);
733                                 return 1;
734                         }
735                         len = sprintf(tmp, "%s/", td->directory);
736                 }
737
738                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
739
740                 for_each_file(td, f, i) {
741                         memset(f, 0, sizeof(*f));
742                         f->fd = -1;
743
744                         if (td->filename)
745                                 sprintf(tmp + len, "%s", td->filename);
746                         else
747                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
748                         f->file_name = strdup(tmp);
749                 }
750         } else {
751                 td->nr_files = 1;
752                 td->files = malloc(sizeof(struct fio_file));
753                 f = &td->files[0];
754
755                 memset(f, 0, sizeof(*f));
756                 f->fd = -1;
757                 f->file_name = strdup(jobname);
758         }
759
760         for_each_file(td, f, i) {
761                 f->file_size = td->total_file_size / td->nr_files;
762                 f->file_offset = td->start_offset;
763         }
764                 
765         fio_sem_init(&td->mutex, 0);
766
767         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
768         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
769         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
770
771         if (td->stonewall && td->thread_number > 1)
772                 groupid++;
773
774         td->groupid = groupid;
775
776         if (setup_rate(td))
777                 goto err;
778
779         if (td->write_lat_log) {
780                 setup_log(&td->ts.slat_log);
781                 setup_log(&td->ts.clat_log);
782         }
783         if (td->write_bw_log)
784                 setup_log(&td->ts.bw_log);
785
786         if (!td->name)
787                 td->name = strdup(jobname);
788
789         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
790
791         if (!terse_output) {
792                 if (!job_add_num) {
793                         if (td->io_ops->flags & FIO_CPUIO)
794                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
795                         else {
796                                 char *c1, *c2, *c3, *c4;
797
798                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
799                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
800                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
801                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
802
803                                 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);
804
805                                 free(c1);
806                                 free(c2);
807                                 free(c3);
808                                 free(c4);
809                         }
810                 } else if (job_add_num == 1)
811                         fprintf(f_out, "...\n");
812         }
813
814         /*
815          * recurse add identical jobs, clear numjobs and stonewall options
816          * as they don't apply to sub-jobs
817          */
818         numjobs = td->numjobs;
819         while (--numjobs) {
820                 struct thread_data *td_new = get_new_job(0, td);
821
822                 if (!td_new)
823                         goto err;
824
825                 td_new->numjobs = 1;
826                 td_new->stonewall = 0;
827                 job_add_num = numjobs - 1;
828
829                 if (add_job(td_new, jobname, job_add_num))
830                         goto err;
831         }
832         return 0;
833 err:
834         put_job(td);
835         return -1;
836 }
837
838 /*
839  * Initialize the various random states we need (random io, block size ranges,
840  * read/write mix, etc).
841  */
842 int init_random_state(struct thread_data *td)
843 {
844         unsigned long seeds[4];
845         int fd, num_maps, blocks, i;
846         struct fio_file *f;
847
848         if (td->io_ops->flags & FIO_CPUIO)
849                 return 0;
850
851         fd = open("/dev/urandom", O_RDONLY);
852         if (fd == -1) {
853                 td_verror(td, errno, "open");
854                 return 1;
855         }
856
857         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
858                 td_verror(td, EIO, "read");
859                 close(fd);
860                 return 1;
861         }
862
863         close(fd);
864
865         os_random_seed(seeds[0], &td->bsrange_state);
866         os_random_seed(seeds[1], &td->verify_state);
867         os_random_seed(seeds[2], &td->rwmix_state);
868
869         if (td->sequential)
870                 return 0;
871
872         if (td->rand_repeatable)
873                 seeds[3] = FIO_RANDSEED * td->thread_number;
874
875         if (!td->norandommap) {
876                 for_each_file(td, f, i) {
877                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
878                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
879                         f->file_map = malloc(num_maps * sizeof(long));
880                         f->num_maps = num_maps;
881                         memset(f->file_map, 0, num_maps * sizeof(long));
882                 }
883         }
884
885         os_random_seed(seeds[3], &td->random_state);
886         return 0;
887 }
888
889 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
890 {
891 #ifdef FIO_HAVE_CPU_AFFINITY
892         unsigned int i;
893
894         CPU_ZERO(&cpumask);
895
896         for (i = 0; i < sizeof(int) * 8; i++) {
897                 if ((1 << i) & cpu)
898                         CPU_SET(i, &cpumask);
899         }
900 #endif
901 }
902
903 static int is_empty_or_comment(char *line)
904 {
905         unsigned int i;
906
907         for (i = 0; i < strlen(line); i++) {
908                 if (line[i] == ';')
909                         return 1;
910                 if (line[i] == '#')
911                         return 1;
912                 if (!isspace(line[i]) && !iscntrl(line[i]))
913                         return 0;
914         }
915
916         return 1;
917 }
918
919 static int str_rw_cb(void *data, const char *mem)
920 {
921         struct thread_data *td = data;
922
923         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
924                 td->ddir = DDIR_READ;
925                 td->sequential = 1;
926                 return 0;
927         } else if (!strncmp(mem, "randread", 8)) {
928                 td->ddir = DDIR_READ;
929                 td->sequential = 0;
930                 return 0;
931         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
932                 td->ddir = DDIR_WRITE;
933                 td->sequential = 1;
934                 return 0;
935         } else if (!strncmp(mem, "randwrite", 9)) {
936                 td->ddir = DDIR_WRITE;
937                 td->sequential = 0;
938                 return 0;
939         } else if (!strncmp(mem, "rw", 2)) {
940                 td->ddir = DDIR_READ;
941                 td->iomix = 1;
942                 td->sequential = 1;
943                 return 0;
944         } else if (!strncmp(mem, "randrw", 6)) {
945                 td->ddir = DDIR_READ;
946                 td->iomix = 1;
947                 td->sequential = 0;
948                 return 0;
949         }
950
951         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
952         return 1;
953 }
954
955 static int str_verify_cb(void *data, const char *mem)
956 {
957         struct thread_data *td = data;
958
959         if (!strncmp(mem, "0", 1)) {
960                 td->verify = VERIFY_NONE;
961                 return 0;
962         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
963                 td->verify = VERIFY_MD5;
964                 return 0;
965         } else if (!strncmp(mem, "crc32", 5)) {
966                 td->verify = VERIFY_CRC32;
967                 return 0;
968         }
969
970         log_err("fio: verify types: md5, crc32\n");
971         return 1;
972 }
973
974 /*
975  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
976  */
977 static char *get_mmap_file(const char *str)
978 {
979         char *p = strstr(str, ":");
980
981         if (!p)
982                 return NULL;
983
984         p++;
985         strip_blank_front(&p);
986         strip_blank_end(p);
987         return strdup(p);
988 }
989
990 static int str_mem_cb(void *data, const char *mem)
991 {
992         struct thread_data *td = data;
993
994         if (!strncmp(mem, "malloc", 6)) {
995                 td->mem_type = MEM_MALLOC;
996                 return 0;
997         } else if (!strncmp(mem, "mmaphuge", 8)) {
998 #ifdef FIO_HAVE_HUGETLB
999                 /*
1000                  * mmaphuge must be appended with the actual file
1001                  */
1002                 td->mmapfile = get_mmap_file(mem);
1003                 if (!td->mmapfile) {
1004                         log_err("fio: mmaphuge:/path/to/file\n");
1005                         return 1;
1006                 }
1007
1008                 td->mem_type = MEM_MMAPHUGE;
1009                 return 0;
1010 #else
1011                 log_err("fio: mmaphuge not available\n");
1012                 return 1;
1013 #endif
1014         } else if (!strncmp(mem, "mmap", 4)) {
1015                 /*
1016                  * Check if the user wants file backed memory. It's ok
1017                  * if there's no file given, we'll just use anon mamp then.
1018                  */
1019                 td->mmapfile = get_mmap_file(mem);
1020                 td->mem_type = MEM_MMAP;
1021                 return 0;
1022         } else if (!strncmp(mem, "shmhuge", 7)) {
1023 #ifdef FIO_HAVE_HUGETLB
1024                 td->mem_type = MEM_SHMHUGE;
1025                 return 0;
1026 #else
1027                 log_err("fio: shmhuge not available\n");
1028                 return 1;
1029 #endif
1030         } else if (!strncmp(mem, "shm", 3)) {
1031                 td->mem_type = MEM_SHM;
1032                 return 0;
1033         }
1034
1035         log_err("fio: mem type: malloc, shm, shmhuge, mmap, mmaphuge\n");
1036         return 1;
1037 }
1038
1039 static int str_ioengine_cb(void *data, const char *str)
1040 {
1041         struct thread_data *td = data;
1042
1043         td->io_ops = load_ioengine(td, str);
1044         if (td->io_ops)
1045                 return 0;
1046
1047         log_err("fio: ioengine= libaio, posixaio, sync, syslet-rw, mmap, sgio, splice, cpu, null\n");
1048         log_err("fio: or specify path to dynamic ioengine module\n");
1049         return 1;
1050 }
1051
1052 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1053 {
1054         mlock_size = *val;
1055         return 0;
1056 }
1057
1058 #ifdef FIO_HAVE_IOPRIO
1059 static int str_prioclass_cb(void *data, unsigned int *val)
1060 {
1061         struct thread_data *td = data;
1062
1063         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1064         return 0;
1065 }
1066
1067 static int str_prio_cb(void *data, unsigned int *val)
1068 {
1069         struct thread_data *td = data;
1070
1071         td->ioprio |= *val;
1072         return 0;
1073 }
1074 #endif
1075
1076 static int str_exitall_cb(void)
1077 {
1078         exitall_on_terminate = 1;
1079         return 0;
1080 }
1081
1082 static int str_cpumask_cb(void *data, unsigned int *val)
1083 {
1084         struct thread_data *td = data;
1085
1086         fill_cpu_mask(td->cpumask, *val);
1087         return 0;
1088 }
1089
1090 /*
1091  * This is our [ini] type file parser.
1092  */
1093 static int parse_jobs_ini(char *file, int stonewall_flag)
1094 {
1095         unsigned int global;
1096         struct thread_data *td;
1097         char *string, *name;
1098         fpos_t off;
1099         FILE *f;
1100         char *p;
1101         int ret = 0, stonewall;
1102
1103         f = fopen(file, "r");
1104         if (!f) {
1105                 perror("fopen job file");
1106                 return 1;
1107         }
1108
1109         string = malloc(4096);
1110         name = malloc(256);
1111         memset(name, 0, 256);
1112
1113         stonewall = stonewall_flag;
1114         do {
1115                 p = fgets(string, 4095, f);
1116                 if (!p)
1117                         break;
1118                 if (is_empty_or_comment(p))
1119                         continue;
1120                 if (sscanf(p, "[%255s]", name) != 1)
1121                         continue;
1122
1123                 global = !strncmp(name, "global", 6);
1124
1125                 name[strlen(name) - 1] = '\0';
1126
1127                 td = get_new_job(global, &def_thread);
1128                 if (!td) {
1129                         ret = 1;
1130                         break;
1131                 }
1132
1133                 /*
1134                  * Seperate multiple job files by a stonewall
1135                  */
1136                 if (!global && stonewall) {
1137                         td->stonewall = stonewall;
1138                         stonewall = 0;
1139                 }
1140
1141                 fgetpos(f, &off);
1142                 while ((p = fgets(string, 4096, f)) != NULL) {
1143                         if (is_empty_or_comment(p))
1144                                 continue;
1145
1146                         strip_blank_front(&p);
1147
1148                         if (p[0] == '[')
1149                                 break;
1150
1151                         strip_blank_end(p);
1152
1153                         fgetpos(f, &off);
1154
1155                         /*
1156                          * Don't break here, continue parsing options so we
1157                          * dump all the bad ones. Makes trial/error fixups
1158                          * easier on the user.
1159                          */
1160                         ret |= parse_option(p, options, td);
1161                 }
1162
1163                 if (!ret) {
1164                         fsetpos(f, &off);
1165                         ret = add_job(td, name, 0);
1166                 } else {
1167                         log_err("fio: job %s dropped\n", name);
1168                         put_job(td);
1169                 }
1170         } while (!ret);
1171
1172         free(string);
1173         free(name);
1174         fclose(f);
1175         return ret;
1176 }
1177
1178 static int fill_def_thread(void)
1179 {
1180         memset(&def_thread, 0, sizeof(def_thread));
1181
1182         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1183                 perror("sched_getaffinity");
1184                 return 1;
1185         }
1186
1187         /*
1188          * fill default options
1189          */
1190         fill_default_options(&def_thread, options);
1191
1192         def_thread.timeout = def_timeout;
1193         def_thread.write_bw_log = write_bw_log;
1194         def_thread.write_lat_log = write_lat_log;
1195
1196 #ifdef FIO_HAVE_DISK_UTIL
1197         def_thread.do_disk_util = 1;
1198 #endif
1199
1200         return 0;
1201 }
1202
1203 static void usage(void)
1204 {
1205         printf("%s\n", fio_version_string);
1206         printf("\t--output\tWrite output to file\n");
1207         printf("\t--timeout\tRuntime in seconds\n");
1208         printf("\t--latency-log\tGenerate per-job latency logs\n");
1209         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1210         printf("\t--minimal\tMinimal (terse) output\n");
1211         printf("\t--version\tPrint version info and exit\n");
1212         printf("\t--help\t\tPrint this page\n");
1213         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1214 }
1215
1216 static int parse_cmd_line(int argc, char *argv[])
1217 {
1218         struct thread_data *td = NULL;
1219         int c, ini_idx = 0, lidx, ret;
1220
1221         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1222                 switch (c) {
1223                 case 't':
1224                         def_timeout = atoi(optarg);
1225                         break;
1226                 case 'l':
1227                         write_lat_log = 1;
1228                         break;
1229                 case 'w':
1230                         write_bw_log = 1;
1231                         break;
1232                 case 'o':
1233                         f_out = fopen(optarg, "w+");
1234                         if (!f_out) {
1235                                 perror("fopen output");
1236                                 exit(1);
1237                         }
1238                         f_err = f_out;
1239                         break;
1240                 case 'm':
1241                         terse_output = 1;
1242                         break;
1243                 case 'h':
1244                         usage();
1245                         exit(0);
1246                 case 'c':
1247                         ret = show_cmd_help(options, optarg);
1248                         exit(ret);
1249                 case 'v':
1250                         printf("%s\n", fio_version_string);
1251                         exit(0);
1252                 case FIO_GETOPT_JOB: {
1253                         const char *opt = long_options[lidx].name;
1254                         char *val = optarg;
1255
1256                         if (!strncmp(opt, "name", 4) && td) {
1257                                 ret = add_job(td, td->name ?: "fio", 0);
1258                                 if (ret) {
1259                                         put_job(td);
1260                                         return 0;
1261                                 }
1262                                 td = NULL;
1263                         }
1264                         if (!td) {
1265                                 int global = !strncmp(val, "global", 6);
1266
1267                                 td = get_new_job(global, &def_thread);
1268                                 if (!td)
1269                                         return 0;
1270                         }
1271
1272                         ret = parse_cmd_option(opt, val, options, td);
1273                         if (ret) {
1274                                 log_err("fio: job dropped\n");
1275                                 put_job(td);
1276                                 td = NULL;
1277                         }
1278                         break;
1279                 }
1280                 default:
1281                         break;
1282                 }
1283         }
1284
1285         if (td) {
1286                 ret = add_job(td, td->name ?: "fio", 0);
1287                 if (ret)
1288                         put_job(td);
1289         }
1290
1291         while (optind < argc) {
1292                 ini_idx++;
1293                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1294                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1295                 optind++;
1296         }
1297
1298         return ini_idx;
1299 }
1300
1301 static void free_shm(void)
1302 {
1303         struct shmid_ds sbuf;
1304
1305         if (threads) {
1306                 shmdt((void *) threads);
1307                 threads = NULL;
1308                 shmctl(shm_id, IPC_RMID, &sbuf);
1309         }
1310 }
1311
1312 /*
1313  * The thread area is shared between the main process and the job
1314  * threads/processes. So setup a shared memory segment that will hold
1315  * all the job info.
1316  */
1317 static int setup_thread_area(void)
1318 {
1319         /*
1320          * 1024 is too much on some machines, scale max_jobs if
1321          * we get a failure that looks like too large a shm segment
1322          */
1323         do {
1324                 size_t size = max_jobs * sizeof(struct thread_data);
1325
1326                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1327                 if (shm_id != -1)
1328                         break;
1329                 if (errno != EINVAL) {
1330                         perror("shmget");
1331                         break;
1332                 }
1333
1334                 max_jobs >>= 1;
1335         } while (max_jobs);
1336
1337         if (shm_id == -1)
1338                 return 1;
1339
1340         threads = shmat(shm_id, NULL, 0);
1341         if (threads == (void *) -1) {
1342                 perror("shmat");
1343                 return 1;
1344         }
1345
1346         atexit(free_shm);
1347         return 0;
1348 }
1349
1350 /*
1351  * Copy the fio options into the long options map, so we mirror
1352  * job and cmd line options.
1353  */
1354 static void dupe_job_options(void)
1355 {
1356         struct fio_option *o;
1357         unsigned int i;
1358
1359         i = 0;
1360         while (long_options[i].name)
1361                 i++;
1362
1363         o = &options[0];
1364         while (o->name) {
1365                 long_options[i].name = o->name;
1366                 long_options[i].val = FIO_GETOPT_JOB;
1367                 if (o->type == FIO_OPT_STR_SET)
1368                         long_options[i].has_arg = no_argument;
1369                 else
1370                         long_options[i].has_arg = required_argument;
1371
1372                 i++;
1373                 o++;
1374                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1375         }
1376 }
1377
1378 int parse_options(int argc, char *argv[])
1379 {
1380         int job_files, i;
1381
1382         f_out = stdout;
1383         f_err = stderr;
1384
1385         options_init(options);
1386
1387         dupe_job_options();
1388
1389         if (setup_thread_area())
1390                 return 1;
1391         if (fill_def_thread())
1392                 return 1;
1393
1394         job_files = parse_cmd_line(argc, argv);
1395
1396         for (i = 0; i < job_files; i++) {
1397                 if (fill_def_thread())
1398                         return 1;
1399                 if (parse_jobs_ini(ini_file[i], i))
1400                         return 1;
1401                 free(ini_file[i]);
1402         }
1403
1404         free(ini_file);
1405
1406         if (!thread_number) {
1407                 log_err("No jobs defined(s)\n");
1408                 return 1;
1409         }
1410
1411         return 0;
1412 }